Download Intermediary Asset Pricing

Document related concepts

Stock market wikipedia , lookup

Financial crisis of 2007–2008 wikipedia , lookup

Securitization wikipedia , lookup

Market sentiment wikipedia , lookup

Financial Crisis Inquiry Commission wikipedia , lookup

Black–Scholes model wikipedia , lookup

Private equity secondary market wikipedia , lookup

Efficient-market hypothesis wikipedia , lookup

Systemically important financial institution wikipedia , lookup

Stock selection criterion wikipedia , lookup

Derivative (finance) wikipedia , lookup

Bridgewater Associates wikipedia , lookup

Hedge (finance) wikipedia , lookup

Investment fund wikipedia , lookup

Financial crisis wikipedia , lookup

Systemic risk wikipedia , lookup

Transcript
Intermediary Asset Pricing
Zhiguo He
Arvind Krishnamurthy
∗
November 6, 2007
Abstract
We present an equilibrium asset pricing model in which intermediaries are marginal in setting prices.
The intermediaries we model are hedge funds, mutual funds, banks, or insurance companies. We calibrate
the model to a hedge fund crisis episode where parameters are chosen so that the marginal investor
resembles a hedge fund with leverage. We are able to qualitatively and quantitatively match the behavior
of risk premia and interest rates in a financial crisis. Moreover, the model captures the slow mobility of
capital during a crisis and can replicate observed crisis recovery times. We also calibrate our model to a
broad intermediation scenario where parameters are chosen so that the marginal investor is an amalgam
of the intermediaries we observe in practice. We show that the intermediation effects help to generate a
volatile pricing kernel and a market risk premium matching the empirically observed equity premium.
JEL Codes: G12, G2, E44
Keywords: Liquidity, LTCM Crisis, Delegation, Financial Institutions, Asset Pricing, Collateral, Credit.
∗ Northwestern
University. E-mails: [email protected], [email protected]. We thank Vadim
Linetsky, Mark Loewenstein, Annette Vissing-Jorgensen, Hongjun Yan and seminar participants at Yale and the FDIC for
helpful comments.
1
1
Introduction
This paper presents an equilibrium asset pricing model in which intermediaries are marginal in setting prices.
The intermediaries we model are most naturally thought of as hedge funds, mutual funds, banks, or insurance
companies. Although over 50% of financial wealth in the U.S. is invested through these types of intermediaries
(Allen, 2001), traditional approaches to asset pricing ignore intermediation by invoking the assumption that
intermediaries’ actions reflect the preferences of their client-investors. With this assumption, the traditional
approach treats intermediaries as a “veil,” and instead posits that a representative household is marginal in
pricing all assets.
We deviate from the traditional approach for two reasons. First, it is by now widely accepted that the
turmoil in financial markets in the fall of 1998 was due to shocks among hedge funds and other sophisticated
intermediaries (see Xiong, 2001, Kyle and Xiong, 2001, Gromb and Vayanos, 2002, Brunnermeier and Pedersen, 2006, He and Krishnamurthy, 2006, or Gabaix, Krishnamurthy, and Vigneron, 2007). Large shocks
to these intermediaries forced them to liquidate positions in several asset markets, driving up risk premia
on these assets. It was intermediaries that were affected rather than a representative household, suggesting
that the marginal investor in the affected asset markets during this episode was an intermediary.1
Second, as is well known, the traditional approach has some shortcomings. In particular, since the
marginal investor is assumed to be a household whose consumption is equal to NIPA aggregate consumption,
variability in the marginal pricing condition for assets – the pricing kernel – is tied to variability in aggregate
consumption growth. However, as the volatility of aggregate consumption growth is low and consumption
growth and asset payoffs are only weakly correlated, it is hard to understand the size of risk premia and
volatility on financial assets under the representative household approach. In our approach, the pricing
kernel for assets is affected by fluctuations in the financial position of intermediaries, which is plausibly an
order of magnitude more volatile than household consumption. Factors such as the volatility of asset values,
the leverage of intermediaries, and the ability of intermediaries to raise capital, all contribute to determining
the intermediary pricing kernel. We show that with a realistic calibration, the model can match empirically
observed risk premia and Sharpe ratios.
1 There
is a growing body of empirical evidence documenting the effects of intermediation constraints (such as capital or
collateral constraints) on asset prices. These studies include, research on mortgage-backed securities (Gabaix, Krishnamurthy,
and Vigneron, 2005), corporate bonds (Collin-Dufresne, Goldstein, and Martin, 2001), default swaps (Berndt, et. al., 2004),
catastrophe insurance (Froot and O’Connell, 1999, 2001), and index options (Bates, 2003; Garleanu, Pedersen, and Poteshman,
2005).
2
The contribution of our paper is to work out an equilibrium model of intermediation that is dynamic,
parsimonious, and can be realistically calibrated. While there is a prior body of work that studies the effect
of intermediation on asset prices, the models employed have almost exclusively been static and designed
to highlight qualitative effects. Allen and Gale in a number of papers show how the financial structure
of intermediaries plays an important role in financial crises (see Allen and Gale, 2005).2 Holmstrom and
Tirole (1997) show how capital constraints in intermediation can affect the equilibrium interest rate as well
as interest rate spreads. Shleifer and Vishny (1997) argue that the tendency for investors to withdraw
funds from intermediaries following negative performance limits the ability of intermediaries to exploit high
returns.3 Our paper draws on the ideas from this prior literature, incorporating these ideas into a fully
dynamic and quantitative general equilibrium model.
Section 2 of the paper presents the model. It consists of a household sector that cannot directly invest
in a risky “intermediated” asset, and an intermediary sector that can invest in the risky asset on behalf of
the households. This modeling captures the idea that the intermediary investment requires some expertise
that only the intermediaries possess. The modeling also borrows from the literature on limited participation
and immediately implies that the household is not marginal in pricing the intermediated asset. The intermediaries, whose investment decisions are taken by a class of agents who we call specialists, are marginal in
pricing the risky asset.
Our key assumption is that the supply of intermediation to households may be constrained. In particular,
we assume that when the specialists (who manage the intermediaries) have low current wealth, households
are reluctant to invest with intermediaries and instead hold their savings in a riskless bond. We offer two
interpretations of the wealth constraint on intermediation. First, we can think of the specialist as the
insiders of a hedge fund, and the specialist’s wealth as the “capital” of the hedge fund. Then states with
low wealth/capital in the model are states where hedge funds are capital constrained (as in Holmstrom and
Tirole, 1997). A second interpretation is that the wealth of a specialist is a summary of the past investment
decisions and realized returns of the specialist. Then we can think of low wealth states as states in which
households pull out of mutual funds because they have delivered low past returns, and mutual funds in turn
2 The
work of Allen and Gale builds on the Diamond and Dybvig (1983) framework. See also Diamond (1997) and Diamond
and Rajan (2005) for models with linkages between the asset market and financial intermediaries.
3 Other papers in the literature on asset pricing and intermediation include Allen and Gorton (1993), Allen and Gale (1994),
Brennan (1993), Grossman and Zhou (1996), Holmstrom and Tirole (1997), Shleifer and Vishny (1997), Dasgupta, Prat and
Verardo (2005), and Vayanos (2005).
3
are forced to liquidate their asset holdings (as in Shleifer and Vishny, 1997). In both cases, the key dynamic
of the model is that low specialist wealth states lead households to withdraw funds from intermediaries and
indirectly reduce their participation in the risky asset market. This dynamic then drives up the risk premium
on the risky asset.
We calibrate our model to two scenarios. As noted above, intermediaries are thought to play an important
role in financial crises. The first scenario we present is a hedge-fund crisis episode where we parameterize the
model so that the marginal investor during a crisis resembles a hedge fund with leverage. This calibration
is explained in Section 4.
The striking feature of financial crises is the sudden and dramatic increase of risk premia. For example, in
the hedge fund crisis of the fall of 1998, many credit spreads and mortgage-backed security spreads doubled
from their pre-crisis levels. Our baseline calibration can replicate this dramatic behavior. We find that when
the intermediation capital constraint does not bind, risk premia are not very sensitive to the state. On
the other hand, when constraints bind, risk premia and Sharpe ratios increase more than linearly with the
tightness of constraints. Simulating the model, we find that the average risk premium conditional on the
capital constraint not binding is 4.9%. The conditional average Sharpe ratio is 40%. Using these numbers
to reflect a pre-crisis normal level, we find that the probability of the risk premium exceeding 7.5% is 4%.
The probability of the risk premium exceeding 10%, which is twice the “normal” level, is 1%. The 1998
episode saw risk premia and Sharpe ratios rise considerably, in the range of 1.5X to 2X. Our model puts
the probability of such an event between 1% and 4%.
Another important feature of financial crises is the pattern of recovery of spreads. In the 1998 crisis, most
spreads took about 10 months to halve from their crisis-peak levels to pre-crisis levels. As we discuss later in
the paper, half-lives of between 6 months and extending over a year have been documented in a variety of asset
markets and crisis situations. We note that these types of recovery patterns are an order of magnitude slower
than the daily mean reversion patterns documented in the market microstructure literature (e.g., Campbell,
Grossman, and Wang, 1993). A common wisdom among many observers is that this recovery reflects the slow
movement of capital into the affected markets (Froot and O’Connell, 1999, Berndt, et. al., 2004, Mitchell,
Pedersen, and Pulvino, 2007). Our baseline calibration of the model can replicate these speeds of capital
movement. In a crisis state, risk premia are high and the specialists hold leveraged positions on the risky
asset. Over time, profits from this position increases the capital base of the intermediaries, thereby relaxing
the intermediation constraint. Households mirror the rise in intermediary capital by increasing the allocation
4
of their own capital to the intermediaries. Together these forces lead to increased risk-bearing capacity and
lower risk premia. In our baseline calibration, we show that simulating the model starting from an extreme
crisis state (risk premium of 20%), the half-life of the risk premium back to the unconditional average risk
premium is about 5.5 months. From a risk premium of 10%, which is about twice the unconditional average
risk premium, the half-life is close to 19 months.
The second exercise we perform with our model is an aggregate asset pricing calibration. At heart our
model is a heterogeneous agent one where, because of a friction, the marginal investor’s consumption is more
volatile than that of the average investor. In the literature, heterogeneous agent models building on a similar
mechanism have been proposed to explain the equity premium puzzle (for example, see Mankiw and Zeldes
(1991) and Vissing-Jorgensen (2002)).4 Thus, it is interesting to see how our model fares in this regard.
We suppose that the risky asset of the model encompasses all risky financial assets including stocks,
mortgages, etc., and that all investment in these assets is made by intermediaries. While clearly in practice
there are investments in risky assets that do not require the expertise of an intermediary, the exercise provides
a benchmark for our intermediation model. As noted above, over 50% of financial assets are held through
intermediaries.
We parameterize the model so that the marginal investor looks like an amalgam of all intermediaries,
including hedge funds, mutual funds, banks and insurance companies. In this scenario, our parameter
choices lead to a more muted effect of the constraint than in the crisis scenario. On the other hand, the
parameters have the economy almost always in a state where the constraint has an effect on the intermediaries’
investments.
Simulating the model, we find that the unconditional average risk premium is 5.63%. This number is
within the range of estimates of the equity risk premium, which is one guide as to the expected return
on risky assets. The Sharpe ratio on the risky asset is 0.44. The risk premium also varies depending on
economic conditions. For example, variation in the intermediation constraint may be driven by changes in
the financial health of the intermediation sector over the business cycle. If we start from the mean state in
the simulation, and move one standard deviation towards a more constrained state the risk premium rises
to 6.38%. Moving one standard deviation towards a less constrained states, the risk premium falls to 4.32%.
This 206 bps variation provides a gauge as to the business cycle variation in risk premia induced by the
4 The
aggregate asset pricing literature on the equity premium puzzle falls into one of two branches. One looks to solve
the puzzle by investigating alternative preference formulations for a representative household (Campbell and Cochrane (1999),
Barberis, Huang, and Santos (2001)). The other branch of the literature studies heterogeneous agent models.
5
model.
From a theoretical standpoint, the intermediation approach of this paper offers an avenue to introduce
liquidity effects into a general equilibrium asset pricing model. In our two asset model, the risky asset is
illiquid in the sense that only a subset of the agents in the economy can directly trade this asset. The riskless
asset is liquid since all agents participate in the market. From this standpoint, the risk premium of our model
is at least partly a liquidity premium. This notion of liquidity, deriving from market segmentation, is most
similar to Allen and Gale (1994). Many of our results can likewise be viewed as liquidity effects. The main
result of the analysis that changes in intermediary capital endogenously affects asset prices traces a connection
between disintermediation, participation, and the liquidity premium on the risky asset. Particularly in the
calibration of scenario 1, we show that when intermediary capital falls low enough the model can replicate
a liquidity crisis event. We also argue that the liquidity factor of Pastor and Stambaugh (2003) and Sadka
(2006) may proxy for intermediary capital. Since the pricing kernel in the model is a function of intermediary
capital, our model helps to understand why liquidity may be a priced factor. Finally, we study an extension
of the model where we introduce a government that maintains a positive supply of the riskless bond. We
show that increases in the supply of the riskless bond, by expanding the supply of liquid assets, lowers the risk
premium on the risky asset. This exercise helps to explain the liquidity effect documented by Krishnamurthy
and Vissing-Jorgensen (2007) that increases in the supply of Treasury bonds lowers the liquidity premium
on other assets relative to Treasury bonds.
The paper is organized as follows. Sections 2 and 3 outline the model and its solution. Section 4 explains
how we calibrate the model. Section 5 presents the results of the hedge-fund crisis calibration. Section 6
presents the aggregate asset pricing calibration. Section 7 studies the extension with government bonds in
positive supply. Section 8 concludes and is followed by an Appendix with further details of the derivation of
the model solution.
2
The Model: Intermediation and Asset Prices
Our model is a variant of a traditional endowment economy, along the lines of the Lucas (1978) tree model.
The economy is infinite-horizon, continuous-time, and has a single perishable consumption good, which we
will use as the numeraire. There are two assets, a riskless bond in zero net supply, and a stock that pays a
risky dividend. We normalize the total supply of stocks to be one unit.
6
The stock pays a dividend of Dt per unit time, where {Dt : 0 ≤ t < ∞} follows a geometric Brownian
motion,
dDt
= gdt + σdZt
Dt
given
D0 .
(1)
g > 0 and σ > 0 are constants. Throughout this paper {Z} = {Zt : 0 ≤ t < ∞} is a standard Brownian
motion on a complete probability space (Ω, F , P) with an augmented filtration {Ft : 0 ≤ t < ∞} generated
by the Brownian motion {Z}. We denote the progressively measurable processes {Pt : 0 ≤ t < ∞} and
{rt : 0 ≤ t < ∞} as the stock price and interest rate processes, respectively. We also define the total return
on the stock as,
dRt =
Dt dt + dPt
.
Pt
(2)
To this standard setting, we introduce heterogeneity among agents and a need for intermediation. There
are two classes of agents in the economy, households and specialists. We assume that the households cannot
invest directly in the risky asset, while the specialists can directly invest in the risky asset. The riskless
asset is available to all agents. While households are restricted from directly investing in the risky asset, we
assume that the specialists manage intermediaries that raise funds from households and invest these funds
in the risky asset on behalf of the households. In our model, the households demand intermediation services
while the specialists supply these services.
The restriction on the households’ investment choices is similar to assumptions made in the literature on
limited market participation (e.g., Mankiw and Zeldes, 1991, Allen and Gale, 1994, Basak and Cuoco, 1998,
Vissing-Jorgensen, 2002). Unlike this literature, we allow the agents who do participate in the market (the
specialists) to invest in the risky asset on behalf of the households. In our context, this modeling captures
the idea that the investment in the risky asset requires some expertise which only the specialists possess.
The investment restriction and scope for intermediation are our principal modeling departures from a
standard consumption asset pricing model. We elaborate on these items in the next subsections.
2.1
Households: The demand for intermediation
We begin by describing the household side of the economy. As will become clear, households are not
marginal in the asset market and thus play a secondary role in determining asset prices. The main role of
the household sector is in determining the amount of savings channeled to the intermediaries. We model
the household sector as an overlapping generation (OG) of agents. This keeps the decision problem of the
household fairly simple. On the other hand, we enrich the model to include household labor income and
7
introduce heterogeneity within the household sector. Both enrichments are useful in realistically calibrating
the model.
For the sake of clarity in explaining the OG environment in a continuous time model, we index time as
t, t + δ, t + 2δ, ... and consider the continuous time limit when δ is of order dt. A unit mass of generation t
agents are born with wealth wth and live in periods t and t + δ. They maximize utility:
h
ρδ ln cht + (1 − ρδ) Et [v(wt+δ
)].
cht is the household’s consumption in period t and v(·) is a bequest function over wealth left for generation
t + δ. We assume that the bequest function is logarithmic, so that the household objective is:
h
ρδ ln cht + (1 − ρδ) Et [ln wt+δ
].
(3)
In addition to wealth of wth , we assume that generation t households receive labor income at date t of
l Dt δ. l > 0 is a constant and Dt is the dividend on the risky asset at time t. Labor income is assumed
proportional to dividends in order to preserve some useful homogeneity properties of the equilibrium. We
introduce labor income to more realistically match the consumption-savings profile of households. Providing
the households some labor income also ensures that the economy never reaches a state where households
“die” out.
It is easy to verify that as δ → dt in the continuous time limit, the household’s consumption rule is,
cht = ρwth .
(4)
In particular, note that the labor income does not affect the consumption rule because the labor income flow
is of order dt. Interpreting ρ > 0 as the household’s rate of time preference, we note that this is the standard
consumption rule for logarithmic agents. The household is “myopic” and his rule does not depend on his
investment opportunity set.
The household invests his wealth from t to t + δ in financial assets. As noted earlier, households are not
directly able to save in the risky asset and can only directly access the riskless bond market. We assume
that the household can choose any positive level of bond holdings when saving in the riskless bond (note
that short-selling of the bond is rule out). The household must use an intermediary when accessing the risky
asset market.
We consider a further degree of heterogeneity in the intermediation investment restriction. We assume
that a fraction λ of the households can ever only invest in the riskless bond. The remaining fraction, 1 − λ,
8
may enter the intermediation market and save a fraction of their wealth with intermediaries which indirectly
invest in the risky asset on their behalf. We refer to the former as “debt households” and the latter as “stock
households.”5
The heterogeneity among households is realistic. Clearly, there are many households that only save in
a bank account. In the literature cited earlier on limited market participation, all households are “debt
households.” The demand for intermediation in our model stems from the stock households. Introducing
this degree of heterogeneity allows for a better model calibration.
To summarize, a debt and stock household are born at generation t with wealth of wth . The households
receive labor income and choose a consumption rate of ρwth . They also make savings decisions, respecting
the restriction on their investment options.
The debt household’s consumption decision, given wealth of wth , is described by (4). The savings decision
is to invest wth in the bond market at interest rate of rt .
The stock household’s consumption is also described by (4). His portfolio decision is how much wealth to
allocate to intermediaries. We denote αht ∈ [0, 1] as the fraction of the household’s wealth in the intermediary
f t . The remaining 1 − αh of household wealth is invested in the
and denote the intermediary’s return as dR
t
riskless bond and earns the interest rate of rt dt. The stock household chooses αht to maximize (3). Given
the log objective function, this decision solves,
f t ].
f t ] − 1 αh 2 V art [dR
max αht Et [dR
2 t
αh
t ∈[0,1]
Our description of the stock household’s problem is not yet complete. In subsection 2.3 we will describe a
constraint on this portfolio problem stemming from the supply of intermediation. For now, we note that
αht (1 − λ)wth of household wealth is saved in intermediaries.
Given the decisions by the debt household and the stock household, the evolution of wth across generations
is described by,
5 The
f t − rt dt .
dwth = (lDt − ρwth )dt + wth rt dt + αht (1 − λ)wth dR
(5)
wealth of the debt household and stock household evolve differently between t and t + δ. We assume that this wealth
is pooled together and distributed equally to all agents of generation t + δ. The latter assumption ensures that we do not need
to keep track of the distribution of wealth over the households when solving for the equilibrium of the economy.
9
2.2
Specialists and intermediation
There is a unit mass of identical specialists who manage the intermediaries in which the households invest.
The specialists represent the insiders/decision-makers of a hedge fund or a mutual fund. We collapse all
of an intermediary’s insiders into a single agent, following the device of modeling entrepreneur-managers of
firms in the corporate finance literature (e.g. Holmstrom and Tirole, 1997).
Formally, we assume that the specialists are infinitely-lived and maximize an objective function,
E
Z
∞
−ρt
e
u(ct )dt
0
ρ > 0;
(6)
where ct is the date t consumption rate of the specialist. We consider a CRRA instantaneous utility function
with parameter γ for the specialists, u(ct ) =
1−γ
1
.
1−γ ct
Note the specialists are infinitely lived while households are modeled using the OG structure. As we will
see, specialists play the key role in determining asset prices. Our modeling ensures that their choices reflect
the forward-looking dynamics of the economy. We treat households in a simpler manner for tractability
reasons. We deem the cost of the simplification to be low since households play a secondary role in the
model.
Each specialist manages one intermediary. We denote the date t wealth of specialists as wt and assume
that this is wholly invested in the intermediary. We think of wt as the specialist’s “stake” in the intermediary,
possibly capturing financial wealth at risk in the intermediary. Although outside the scope of the model,
we may imagine that wt also captures reputation that is at stake in the intermediary and the future income
from being an insider of the intermediary. Specialists do not receive additional labor income.
We envision the following to describe the interaction between specialists and households. At every t, each
specialist is randomly matched with a stock household to form an intermediary. These interactions occur
instantaneously and result in a continuum of (identical) bilateral relationships.6 The household allocates the
desired funds of αht (1 − λ)wth to the intermediary. Specialists then execute trades for the intermediary in a
Walrasian stock and bond market, and the household trades in only the bond market. At t + dt the match
is broken, and the intermediation market repeats itself.
Consider one of the intermediary relationships between specialist and stock household. The specialist
manages an intermediary whose total assets under management are the sum of the specialist’s wealth, wt ,
6 Why
the matching structure instead of a Walrasian intermediation market? In the Walrasian case, when intermediation is
supply constrained, specialists charge the households a fee for managing the intermediary that depends on the tightness of the
intermediation constraint. In the matching structure the fee is always zero which makes solving the model somewhat easier.
10
and the wealth that the stock household allocates to the intermediary, αht (1 − λ)wth . The specialist makes all
investment decisions on these funds and faces no portfolio restrictions in buying or short-selling either the
risky asset or the riskless bond. Suppose that the specialist chooses to invest a fraction αIt of the portfolio
in the risky asset and 1 − αIt in the riskless asset. Then, the return delivered by the intermediary is,
gt = rt dt + αI (dRt − rt dt),
dR
t
(7)
where dRt is the total return on the risky asset.
2.3
Intermediation constraint
The key assumption of our model is that the household is unwilling to invest more than mwt of funds in
the intermediary (m > 0 is a constant). That is, if the specialist has one dollar of wealth invested in the
intermediary, the household will only invest up to m dollars of his own wealth in the intermediary. He and
Krishnamurthy (2006) derive this sort of capital constraint by assuming moral hazard by the specialist. In
their model, the household requires that the specialist have a sufficient stake in the intermediary to prevent
shirking.7 Here we adopt the constraint in reduced form.
The wealth requirement implies that the supply of intermediation facing a household is at most,
αht (1 − λ)wth ≤ mwt .
(8)
If either m is small or wt is small, the household’s ability to indirectly participate in the risky asset market
will be restricted.
We may interpret the wealth requirement in two ways. First, as noted above, we can think of wt as
the specialist’s stake in the intermediary, and this stake must be sufficiently high for household’s to feel
comfortable with their investment in the intermediary. Thus, one interpretation is that wt reflects the
capital base of a hedge fund. The managers of a hedge fund typically have some of their wealth tied up
in the investments of the hedge fund. Such an arrangement ensures that the incentives of the hedge fund’s
managers and investors are aligned. However, if a hedge fund loses a lot of money then the capital of the
7 The
He and Krishnamurthy (2006) model is adapted from Holmstrom and Tirole (1997). One feature of the contract
derivation in He and Krishnamurthy worth noting is that it assumes that the household cannot observe and dictate the
specialist’s portfolio choice within the intermediary. Without this assumption, it is possible that the household will bribe the
specialist to make a particular portfolio choice. Of course, such a strategy violates the spirit of the model: namely that only the
specialist has the expertise to invest in the risky asset. In this paper, we also make the assumption that the specialist cannot
observe and dictate the specialist’s portfolio choice within the intermediary.
11
hedge fund will be depleted. In this case, investors will be reluctant to contribute money to the hedge fund,
fearing mis-management or further losses. A hedge fund “capital shock” is one phenomena that we can
capture with our model.
Another interpretation, which is more in keeping with regularities in the mutual fund industry, is that the
wealth of a specialist summarizes his past success in making investment decisions. Low wealth then reflects
poor past performance by a mutual fund, which makes households reluctant to delegate investment decisions
to the specialist. The relation between past performance and mutual fund flows is a well-documented
empirical regularity. As wt falls, reflecting poor past performance, investors may withdraw funds from the
mutual fund. Shleifer and Vishny (1997) present a model with a similar feature: the supply of funds to
arbitrageurs in their model is a function of the previous period’s return by the arbitrageur.
Since we adopt constraint (8) in reduced form, we do not take a stand on the interpretation of the
constraint. Indeed, in one of the calibration scenarios, we match the specialist-intermediary to the entire
intermediary sector – including hedge funds, banks, and mutual funds. From this standpoint, it is useful
that the constraint may be appropriate across a variety of intermediaries.
The novel feature of our model is that wt , and the supply of intermediation, evolve endogenously as a
function of shocks and the past decisions of specialists and households. In both the hedge fund and the
mutual fund example, if the intermediation constraint (8) binds, a fall in wt causes households to reduce
their allocation of funds to intermediaries and invest in the riskless bond. Of course, the risky asset still
has to be held in equilibrium. As households indirectly reduce their exposure to the risky asset, via market
clearing, the specialist increases his exposure to the risky asset. To induce the specialist to absorb more risk,
the risky asset price falls and its expected return rises. This dynamic effect of wt on the equilibrium is the
central driving force of our model. We think it arises naturally when considering the equilibrium effects of
intermediation.
gt (see (7)) on their contributions to
We note that both the household and specialist receive the return dR
the intermediary; that is, both household and specialist invest in the equity of the intermediary. Constraint
(8) limits the equity contribution by the household to the intermediary as a function of the specialist’s equity
contribution. It is important to point out that this constraint is not the usual constraint in the literature
on corporate investment and credit rationing (see as an example, Holmstrom and Tirole, 1997). In that
literature, firms face a restriction on the quantity of funds they can borrow using either equity securities or
debt securities. Constraint (8) in our model does not restrict the amount of debt issued by an intermediary
12
and therefore does not restrict the total funds that an intermediary can raise. Intermediaries can short bonds
in our model in the Walrasian bond market. There is no default on such debt contracts in our continuous
time model. Constraint (8) restricts how risk is shared between specialists and households. It is the dynamics
of risk sharing that drives the behavior of asset prices in our model.8
We now complete our description of the stock household’s portfolio problem accounting for the supply
constraint. The stock household solves:
f t ] − 1 αh 2 V art [dR
f t]
max αht Et [dR
t
2
αh
t ∈[0,1]
s.t. αht (1 − λ)wth ≤ mwt .
(9)
Finally, we turn to the decision problem of the specialist. The specialist chooses his consumption rate
and the portfolio decision of the intermediary to solve,
max E
{ct ,αIt }
Z
∞
−ρt
e
0
u(ct ) dt
f t αI − rt dt .
s.t. dwt = −ct dt + wt rt dt + wt dR
t
(10)
We can also rewrite the budget constraint in terms of the underlying return:
dwt = −ct dt + wt rt dt + αIt wt (dRt − rt dt) .
Note that αIt is effectively the specialist’s portfolio share in the risky asset.
2.4
Equilibrium
Definition 1 An equilibrium is a set progressively measurable price processes {Pt } and {rt }, and decisions
{ct , cht , αIt , αht } such that,
1. Given the price processes, decisions solve the consumption-savings problems of the debt household, the
stock household (9) and the specialist (10);
2. Decisions satisfy the intermediation constraint of (8);
3. The stock market clears:
αIt (wt + αht (1 − λ)wth )
= 1;
Pt
8 In
(11)
practice, hedge funds have a great deal of flexibility in using the repo market to borrow funds via debt contracts. They
also borrow from investors via equity contracts, and it is restrictions on the latter that we model. Mutual funds, on the other
hand, typically do not borrow using repo contracts. We have chosen not to introduce an additional degree of heterogeneity
among specialists distinguishing those who borrow and those who do not.
13
4. The goods market clears:
ct + cht = Dt (1 + l).
(12)
Given market clearing in stock and goods markets, the bond market clears by Walras’ law. The market
clearing condition for the stock market reflects that the intermediary is the only direct holder of stocks and
has total funds under management of wt + αht (1 − λ)wth , and the total holding of stock by the intermediary
must equal the supply of stocks.
Finally, an equilibrium relation that proves useful when deriving the solution is that,
wt + wth = Pt .
That is, since bonds are in zero net supply, the wealth of specialists and households must sum to the value
of the risky asset.
2.5
Example
Before diving into the mathematical details of the solution, it is worth going through an example to clarify
the effect of constraint (8), which is the main novelty of our model relative to a standard model.
Suppose that m = 1 and λ = 0. Moreover, suppose we are in a state where wt = 100 and wth = 200.
Then it is clear that since mwt < wth , this is a state where intermediation is constrained by (8). Since the
riskless asset is in zero net supply, the value of the risky asset is equal to the sum of wt and wth (i.e. 300).
Suppose that households saturate the intermediation constraint by investing 100 in intermedaries. Then
intermediaries have total equity contributions of 200 (the households’ 100 plus the specialists’ wt ). Since
intermediaries hold all of the risky asset worth 300, their portfolio share in the risky asset must be equal
to 150%. Their portfolio share in the bond is −50%. That is, the intermediary holds a levered position in
the risky asset. The household’s portfolio shares are 0.5 × 150% = 75% in risky asset; and, 25% in debt.
The households and specialists have different portfolio exposures to the risky asset. But since the specialist
drives the pricing of the risky asset, risk premia must adjust to make the 150% portfolio share optimal.
From this situation, suppose that dividends on the risky asset fall. Then, since the specialists are more
exposed to the risky asset than households, wt falls relative to wth . The shock then further tightens the
intermediation constraint, which can create an amplified response to the shock.
Contrast this situation with one in which there is no intermediation constraint. Suppose that households
invest all of their wealth with the intermediaries. Since intermediaries now have 300 and the risky asset is
14
worth 300, the portfolio share of both specialists and households is equal to 100%.
3
Solution
We derive the equilibrium by conjecturing a candidate pricing function and price process and then solving
agents’ decision problem given these prices. We then verify that given agent decisions, market clearing
conditions recover the conjectured pricing function and price process.
The next subsections outline the main steps in deriving the solution. We present detailed derivations of
these steps in Appendix A.
3.1
State variables and candidate price functions
We look for a stationary Markov equilibrium where the state variables are (yt , Dt ), where yt ≡
wth
Dt
is the
dividend scaled wealth of the household. Our economy with only specialists is a standard CRRA/GBM
economy that has been fully analyzed in the literature. It is well known that in that setting, the only state
variable is Dt and the economy scales linearly with Dt . We thereby guess that in our problem Dt is a state
variable, and that our economy is homogeneous in Dt .
Intermediation frictions imply that the distribution of wealth between households and specialists affects
equilibrium. For example, whether constraint (8) binds or not depends on the relative wealth of households
and specialists. We have some freedom in choosing how to define the wealth distribution state variable. It
turns out that using the scaled households’ wealth y is convenient for the analysis.
We conjecture that the equilibrium evolution of yt may be written as an Ito process which solves the
following Stochastic Differential Equation,
dyt = µy dt + σy dZt .
(13)
We will derive expressions for µy and σy .
Due to the homogeneity of the economy with respect to dividends, we conjecture that the equilibrium
stock price is,
Pt = Dt F (yt )
(14)
where F : R → R is twice continuously differentiable on its relevant domain. F (y) is the price/dividend
ratio of the stock.
We derive relations for the three unknown functions, F (y), µy and σy .
15
3.2
Marginal investor and specialist consumption
While the household faces investment restrictions on his portfolio choices, the specialist (intermediary) is
unconstrained in his portfolio choices. This is an important observation about our model because it implies
that the specialist is always the marginal investor in determining asset prices, while the household may
not be. Standard arguments then tell us that we can express the pricing kernel in terms of the specialist’s
equilibrium consumption process.
We have noted in (4) that the household’s optimal consumption given wth is cht = ρwth , which we can
rewrite as cht = ρyt Dt . Now the market clearing condition for goods (from (12)) is,
ct + ρyt Dt = Dt (1 + l).
Thus, in equilibrium, the specialist consumes:
ct = Dt (1 + l − ρyt ).
(15)
We thereby express specialist consumption as a function of the state variables Dt and yt .
Optimality for the specialist gives us the standard consumption-based asset pricing relations (Euler
equation):9
−ρdt − γEt
dct
1
dct
dct
+ γ(γ + 1)V art
+ Et [dRt ] = γCovt
, dRt
ct
2
ct
ct
We apply Ito’s Lemma to (15) to write
dct
ct
(16)
as a function of µy and σy . We can also apply Ito’s Lemma to
(14) to express dRt as a function of the derivatives of F (y) and the unknown drift and diffusion of yt . Then
we arrive at a differential equation that must be satisfied by µy , σy , and F (y):
Proposition 1 The equilibrium Price/Dividend ratio F (y) satisfies the ordinary differential equation (ODE),
g+
F0
1 F 00 2
µy +
σ
F
2 F y
+
1
F0
γρh
+
σy σ = ρ + γg −
(µy + σy σ)
F
F
1 − ρh y
2
ρh
F0
1
ρh
+γ σ −
σ
σ
+
σ
−
γ(γ
+
1)
σ
−
σ
y
y
y
1 − ρh y
F
2
1 − ρh y
Proof: See Appendix.
9 The
Euler equation is a necessary condition for optimality. In Appendix B, we prove sufficiency.
16
(17)
3.3
Dynamics of household wealth
µy and σy are unknown functions in the ODE and describe the dynamics of the households’ scaled wealth
along the equilibrium path.
Recall that we have previously derived the wealth dynamics of the household in (5), which we reproduce
below as:
f t − rt dt .
dwth = (lDt − ρwth )dt + wth rt dt + αht (1 − λ)wth dR
Wealth dynamics are a function of the portfolio choice αht as well as the interest rate and the return on
intermediaries. Since the intermediary, in turn, holds αIt of the risky asset, we rewrite the wealth dynamics
as function of the primitive return dRt :
dwth = (lDt − ρwth )dt + wth rt dt + αht αIt (1 − λ)wth (dRt − rt dt) .
We noted above that dRt can be expressed in terms of F (y), µy and σy . The interest rate, rt , can also be
derived from the specialist’s Euler equation. Since the price of a short-term bond is always one, the Euler
equation gives,
rt dt = ρdt + γEt
dct
γ(γ + 1)
dct
V art
−
.
ct
2
ct
(18)
The only variable that remains to complete the description of wealth dynamics is αht αIt (1 − λ), which is
the household sector’s exposure to the risky asset return.
We first note that in any state where the intermediation constraint of equation (8) binds, the household
chooses,
αht (1 − λ)wth = mwt .
That is, the binding constraint pins down the household’s portfolio share in the intermediary. To find the
household’s exposure to the risky asset we only need to solve for the fraction of the intermediary’s portfolio
in the risky asset, αIt . Since all stocks are held through the intermediary, the equilibrium market clearing
condition (11) gives,
αIt (wt + mwt )
=1
Pt
We rewrite this expression using the fact that wt + wth = Pt to find,
αI,const
=
t
1
F (y)
1 + m F (y) − y
(19)
Equation (19) reveals an important property of the model: shocks that reduce F (y) increase αIt leading the
intermediary to hold a more risky portfolio. That is, if asset values fall, the intermediation constraint tightens
17
and causes households to reduce wealth allocated to intermediaries. In equilibrium, the intermediaries still
hold the risky asset. They do this by increasing their borrowing and holding a larger fraction of the risky
asset in their portfolio.
As an intermediary’s portfolio becomes more risky, the specialist, who has all of his wealth exposed to the
intermediary’s return, owns a more risky portfolio. However since the specialist makes the intermediary’s
portfolio choices, equilibrium prices must be such that the specialist is induced to choose the more risky
portfolio. As we will see in the next sections, this factor drives up risk premia and influences the determination
of asset prices in the model.
When the intermediation constraint does not bind, the household is unconstrained in choosing αht . We
make an assumption that ensures that αht = 1 in this case:
Parameter Assumption 1 We focus on parameters of the model such that in the absence of any portfolio
restrictions, the stock household will choose to have at least 100% of his wealth invested in the intermediary.10
Under this parameter restriction, the stock household allocates all of his wealth to the intermediary (αht = 1)
when intermediation is not constrained. Recall that as we assume that the household cannot short bonds,
the household cannot allocate more than 100% of his wealth to the intermediary.
We now characterize the conditions under which the intermediation constraint binds. Setting αht = 1 in
(8) yields that the constraint binds when,
(1 − λ)wth ≥ mwt .
Using the knowledge that wth + wt = Pt , we rewrite the inequality to find an expression that gives a cutoff
for the constrained states:
yc =
m
F (yc ).
1+m−λ
This equation has a unique solution in all of our parameterizations. For y > yc , intermediation is constrained
by the specialist’s wealth. Thus in equation (19) we note that when y is larger, the intermediaries’ portfolio
share in the risky asset is also larger, increasing the risk premium effect we highlighted above. In our
numerical solutions, we find that over most of the state space F (·) is decreasing in y. Thus, the effect of y
on αIt is reinforced through the effect of y on F (y).
10 Although
we are unable to provide a precise mathematical condition for this parameter restriction, in our calibration it
appears that γ > 1 is a sufficient condition. Loosely speaking, if γ > 1 the specialist is more risk averse than the household,
suggesting that the household will hold more risky assets than the specialist. But given market clearing in the stock market,
the specialist always holds more than 100% of his wealth in the risky asset.
18
When y < yc , the household is unconstrained in allocating funds to the intermediary. The stock household
chooses αht = 1. Using the market clearing condition for stocks, we find,
αI,unconst
=
t
F (y)
.
F (y) − λy
To summarize, we have expressed αIt , the household’s equilibrium exposure to the risky asset, as a
function of y and F . Then it is straightforward to substitute back into the equation for household wealth
dynamics, (5), to find expressions for µy and σy as a function of y and F . Substituting back into the ODE
in Proposition 1, this gives a final ODE to solve for F (y). Further mathematical details of the derivation
are provided in the Appendix.
3.4
Boundary condition
The model has a natural upper boundary condition on y that is determined by the goods market clearing
condition. Since
ct = Dt (1 + l − ρyt ) ,
and the specialist’s consumption ct must be positive, yt has to be bounded by
yb ≡
1+l
.
ρ
In Appendix B, we show that yb is an entrance-no-exit boundary, and that yt never reaches yb .
On an equilibrium path in which y approaches yb , the specialist’s equilibrium consumption c goes to zero.
Since the specialist’s wealth is D (F (y) − y), one natural guess for the boundary condition at this singular
point yb is
F yb = yb .
(20)
In words, when the specialist’s scaled consumption approaches zero, his scaled wealth (F (y) − y) also
converges to zero. In the argument for verification of optimality of the specialist’s equilibrium strategy
which is detailed in Appendix B, we see that this condition translates to the transversality condition for the
specialist’s budget equation. Therefore the boundary condition (20) is sufficient for the equilibrium presented
in this paper to be well-defined.
19
4
Calibration
4.1
m and λ
We model the intermediation sector’s pricing of assets, and how constraints in intermediation affect asset
prices. Table 1 provides data on the main intermediaries in the US economy. Households hold wealth through
a variety of intermediaries. The numbers suggest that the main intermediaries are banks, retirement funds,
mutual funds, and hedge funds.11 This subsection explains how we think about mapping the institutions of
Table 1 into our model, and how we calibrate m and λ.
Table 1: Intermediation Dataa
Group
Assetsb
Commercial Banks
9,156
Savings & Loans
1,749
Property & Casualty Insurance
1,242
Life Insurance
4,351
Private Pensions
4,527
State & Local Ret Funds
2,661
Federal Ret Funds
1,037
Mutual Funds (excluding Money Funds)
5,882
Closed End Funds and ETFs
273
Hedge Funds
3,406
Debt
8,240
1,670
803
4,076
0
0
0
0
0
2,433
Leverage
0.90
0.95
0.65
0.94
0.00
0.00
0.00
0.00
0.00
0.71
a
Most data is from the Flow of Funds Q3 2005 Levels Tables. The Hedge Fund data is based on
an estimate of total hedge fund capital of $973 billion from Fung and Hsieh (2006) and an estimate
that the average fund leverages up its capital base 3.5 times (taken from McGuire, Remolona and
Tsatsaronis (2005))
b
Assets and Debt are in billions of Dollars
Our model treats the entire intermediary sector as a group of identical institutions, while it is clear from
Table 1 that there is heterogeneity across the modes of intermediation. The model takes a broad-brush
approach at the effects of intermediation on asset prices. One aspect of intermediary heterogeneity which is
worth discussing further is that some of the intermediaries have no debt positions and never take on debt
while other do take on debt (see the last column in Table 1).
In our model, when the intermediation constraint (8) binds, losses among intermediaries lead households
to reduce their equity exposure to these intermediaries. If the intermediaries scale down their asset holdings
proportionately, the asset market will not clear – i.e. the intermediary sector’s assets still have to be held in
equilibrium. In our model, the equilibrium is one where the [identical] intermediaries take on debt and hold
a riskier position in the asset. In practice, if households withdraw money from mutual funds, then mutual
11 We
need to be careful in interpreting these numbers because there is some amount of double counting – i.e. pension funds
invest in hedge funds.
20
funds don’t take on debt. Rather, they reduce their holdings of financial assets and some other entity buys
their financial assets. The other entity may be a hedge fund that temporarily provides liquidity to the mutual
fund, or it may be another mutual fund that buys the liquidated assets. If the buyers have difficulty raising
equity to fund the purchase – e.g., they themselves have suffered losses and had withdrawals – then they
have to raise the funds via debt. In practice, such buyers will likely be hedge funds who temporarily increase
leverage rather than a mutual fund that does not operate through leverage. Thus, mapping our model to
practice, we see that heterogeneity plays a role in dictating who among the intermediary sector increases
their exposure to the risky asset during a period of liquidation. However, we note that when (8) binds, the
marginal investor, both in the model and practice, prices assets based on concentrated risk exposure/leverage
considerations. In this sense, our model captures the marginal investor’s preferences well despite omitting
heterogeneity.12
The preceding discussion highlights the mapping between the intermediation constraint of the model
and constraints in the world. Our choice of the intermediation multiplier m parameterizes the intermediation constraint in our model. We note that m has two effects on the model. First, m determines where
intermediation constraints bind in the state space. That is, we found earlier that the constraint binds when,
yc =
m
F (yc ).
1+m−λ
As F (yc ) is decreasing in yc , a larger m leads to a larger yc thereby making intermediation constraints less
likely to bind. The second effect of m is that, conditional on intermediation constraints binding, a higher m
leads to a greater sensitivity of intermediated funds to the specialist’s stake. That is,
αht (1 − λ)wth = mwt ,
so that a one dollar fall in wt leads to an m dollars fall in intermediated funds.
What constitutes a reasonable value of m across the many modes of intermediation represented in our
model? On the one hand, if we focus on hedge funds, then a large value of m seems appropriate. The only
major hedge fund crisis we have witnessed is during the fall of 1998; and, the crisis during that period was
12 It
is worth pausing and also considering the effect of debt constraints on the dynamic we describe. In 1998, when LTCM
ran into trouble, their credit lines were reduced, preventing them from increasing leverage to hold risky assets. Of course, their
risky assets still had to be held by someone in equilibrium. In practice, the investment banking community and trading desks
absorbed these assets; risk exposures became more concentrated and levered in the few players that remained, exactly as in our
model. In 1998, AIG and Warren Buffet offered to buy LTCM’s portfolio. Presumably their bid prices were dictated by the
next best bid for the assets represented by the investment banking community.
21
dramatic. On the other hand, if we think of intermediation more broadly, mutual fund flows are moderately
sensitive to performance. The effect is always present, and not just during extreme events, suggesting that
yc and m are low.
SCENARIO 1: Hedge Fund Crisis
We calibrate our model to two scenarios based on the preceding discussion. In one scenario, we choose m
to be 4 and 6. With this choice of m, the economy spends most of the time (59 − 68%) in the unconstrained
region, but once in the constrained region, leverage rises quickly. In the constrained region, we interpret
the marginal investor as being a hedge fund and choose m to match parameters of hedge funds. We note
that m also measures the specialist’s inside stake in the intermediary relative to the household’s. Hedge
fund contracts typically pay the manager 20% of the fund’s return in excess of a benchmark, plus 1 − 2%
of funds under management (Fung and Hsieh, 2006). A value of m = 4 implies that the specialist’s inside
stake is 1/5 = 20%. The 20% is an option contract so it is not a full equity stake. The 1% is on funds under
management and therefore grows as the fund is succesful and garners more inflows. Thus, a 20% stake is in
the range of parameters that may reasonably capture a hedge fund manager’s inside stake. We also show an
m = 6 case to provide a sense as to the sensitivity of the results to the choice of m.
Our choice of m affects the dynamics of leverage and risk concentration conditional on being in the
constrained region. In practice, we can see from Table 1 that the intermediary sector always has some
leverage. m does not directly affect the leverage in the unconstrained region.
We choose the parameter λ to match leverage in the unconstrained region. In the unconstrained region,
we interpret the marginal investor as being an amalgam of all intermediaries. Within the model, when λ > 0
some households only demand debt, and the intermediaries supply the debt and thereby achieve leverage
even when intermediation is not constrained.
Across all of the intermediaries of Table 1, the Total Debt/Total Assets ratio is 0.50. However, Banks and
Savings & Loans do not only hold traded assets, which are the subject of our model; they hold non-traded
assets (e.g., commercial loans) as well as traded assets (e.g., mortgage backed securities). If we say that
Banks and Savings & Loans have 50% of their assets in traded or securitized assets and compute aggregate
leverage based on 50% of the debt and assets of Banks and Savings & Loans, the aggregate leverage is 0.43.
If we exclude Banks and Savings & Loans completely, the aggregate leverage is 0.31.
22
In our simulations, we consider λ = 0.5 when we consider the m = 4, 6 cases. This value of λ produces
leverage in the unconstrained region around 0.42, and an unconditional average leverage ratio around 0.49.
Scenario 1, based on hedge funds, is used to replicate asset market behavior around a financial crises,
such as that of the fall of 1998. The second scenario in our calibration is a broad intermediation scenario
where we ask how well our model can explain aggregate asset market measures such as the equity premium.
SCENARIO 2: Broad Intermediation and Aggregate Asset Prices
In the second scenario, we calibrate the model so that the intermediation constraint almost always affects
asset prices, but in a less severe manner than the hedge fund scenario. For this scenario, we envision the asset
of the model to encompass a large class of assets held through intermediaries, and envision the intermediaries
of the model to include mutual funds, banks, etc. We suppose that across this large class of intermediaries,
there is always some feedback between the returns of the intermediaries and the funds that households
allocate to intermediaries.
We choose m to be 1 and 0.5; therefore, substantially smaller than the m for the hedge fund case. When
m = 1, the economy spends 98% of the time in the constrained region. We consider the performance-flow
relation among mutual funds to arrive at m = 1. When the intermediation constraint binds, m measures
the sensitivity of the household’s fund contribution to innovations in wt . The stock household’s contribution
to the intermediary is equal to mwt ; normalized by total assets of the intermediary, the stock household
contributes
m
%
1+m
of funds to the intermediary. Then, a 1% fall in wt leads to a
m
%
1+m
fall in the household’s
contribution, relative to total assets under management of the intermediary. With m = 1, this calculation
translates to 0.5% sensitivity of the household contribution. When m = 0.5, the sensitivity is 0.33%. Warther
(1995) documents that a 1% fall in the aggregate market equity return is correlated with a 0.2% outflow of
funds from equity mutual funds. Remolona, Kleiman, and Gruenstein (1997) present similar mutual fund
evidence in extreme events (Table 7 of the paper). They document that the stock market crash of 1987 led
to a fall of 37.7% in net asset values of growth stock funds and a fund outflow of 4.6%, giving a ratio of 0.12.
They also document that the fallout in the junk bond market in 1989 led to a fall of 1.6% in net asset values
of high yield bond funds and a fund outflow of 2.9% (ratio of 1.8). Our 0.33% to 0.5% numbers are in the
range of these measurements.
We choose a lower value of λ = 0.4 when we consider the m = 0.5, 1 cases. Since in these cases the
23
economy spends almost all of the time in the constrained region and the effect of the constraint is to raise
leverage in the constrained region, choosing a smaller value of λ = 0.4 produces an unconditional average
leverage ratio of 0.46.
4.2
σ and g
The asset payoffs we price are ones where investment requires some expertise. To be concrete, these intermediated payoffs may stem from mutual-fund/hedge-fund trading of individual stocks, mutual-fund/hedgefund/bank investments in mortgage-backed securities, mutual-fund/hedge-fund/bank/insurance company
investments in corporate debt or credit derivatives, hedge-fund trading of portfolios of stocks based on
statistical analysis, or intermediaries’ provision of short-term liquidity to financial markets.
We use the aggregate stock market to benchmark this amalgam of payoffs and set σ = 12% and g = 1.84%.
These values are fairly standard, e.g. see Barberis, Huang, and Santos (2001). In the aggregate asset pricing
exercise of Scenario 2, benchmarking to the stock market seems appropriate. In the hedge fund case of
Scenario 1, the stock market is a less obvious benchmark. As another benchmark, Chan, et. al. (2005)
report the volatility of returns on different categories of hedge funds, finding standard deviations ranging
between 3% to 17%. They also note that these numbers underestimate the true volatility of returns because
the underlying assets of hedge funds are illiquid and because there is evidence that hedge funds smooth
reported returns. The choice of σ = 12% produces an equilibrium return volatility in our model between
12% and 13%.
We note that our choice of σ = 12% is an order of magnitude higher than aggregate consumption
volatility of close to 3%. In standard general equilibrium approaches to asset pricing, exemplified by Campbell
and Cochrane (1999) or Barberis, Huang, and Santos (2001), models assume a representative agent whose
consumption is equal to NIPA aggregate consumption and price a payoff with a dividend stream that matches
properties of aggregate stock market dividends.
The marginal investor in our model is the specialist-intermediary rather than a representative agent
because intermediaries are not a veil. As our analysis shows, the specialist’s marginal utility is endogenously
affected by fluctuations in the value of assets that the specialist holds. Thus, we do not exogenously specify
the marginal investor’s consumption process based on aggregate consumption, but endogenously derive the
joint behavior of specialist consumption and the prices of intermediated assets. For this reason, we choose
the volatility of the risky asset’s dividends to match those of financial payoffs rather than that of aggregate
24
consumption. Indeed, we see the endogenous relationship between financial wealth fluctuations and the
pricing kernel as an important reason to model intermediaries rather than treat them as a veil.
Finally, in principle it seems possible to reconcile the low aggregate consumption volatility we observe
in practice with the 12% dividend volatility of the model by assuming that the household sector’s labor
income is weakly correlated with dividends (as in the data). Unfortunately, such a model will no longer be
homogeneous with respect to dividends which will considerably complicate the analysis.
Table 2: Parameters
Panel A: Intermediation
Scenario 1
m Intermediation multiplier
4, 6
λ
Debt ratio
0.5
Panel B: Preferences and Cashflows
g
Dividend growth
1.84%
σ
Dividend volatility
12%
ρ
Time discount rate
8%
γ
RRA of specialist
2
l
Household labor income ratio
1
4.3
Scenario 2
0.5,1
0.4
SAME
l, γ, and ρ
We choose l to match the income profile of typical household. In our model, households receive expected
h
i
f t − rt dt
capital income of E wt rt dt + (1 − λ)αht dR
and expected labor income of E[lDt dt]. In NIPA
data, capital income as a share of GDP is about 30%. Malloy, Moskowitz, and Vissing-Jorgenson (2006)
report that for the top one-third of households in terms of wealth, the share of capital income in total income
in 2001 was 34%. We choose l based on these considerations. We set l = 1 which produces a capital income
to total income share around 35%.
We choose γ = 2 as risk aversion of the specialist. As noted earlier, the household has logarithmic
preferences. Allowing for γ > 1 for the specialist allows us to capture dynamic hedging effects that would
be absent if we set γ = 1. We choose ρ to match an average riskless interest rate of between 0% and 1%.
This leads us to a value of ρ equal to 0.08. These numbers are all typical in the literature. Finally, our
2
lγρ
parameter choices are also dictated by the restriction that, ρ + g(γ − 1) − γ(γ−1)σ
− 1+l
> 0. This restriction
2
is necessary to ensure that the economy is well-behaved at t = ∞.
25
4.4
Numerical method
We are able to find a closed form solution of the ODE only for the case where γ = 1 (the log case for which
F (·) =
1+l
).
ρ
Rather than restricting attention to that special case, we present numerical solutions based on
the calibration of Table 2. We use one of MATLAB’s built-in ODE solvers to derive solutions for F (y), µy ,
and σy . Further details are provided in the Appendix.
With these solutions in hand, we numerically simulate the model to obtain the steady state distribution
of the state variable y as well as a number of asset price measurements that we report in the next sections.
We begin the economy at a state (y0 = yc , D0 = 1) and simulate the economy for 5000 years. That is
we obtain a sequence of independent draws from the normal distribution and use these draws to represent
innovations in our shock process Zt . The path of Zt can then be mapped into a path of the state variable.
We compute the time-series averages of a number of relevant asset price measurements from years 1000 to
5000 of this sample. The simulation unit is monthly, and based on those monthly observations we compute
annual averages. We repeat this exercise 5000 times, averaging across all of the simulated Zt paths. We
find that changing the starting value y0 does not affect the computed distribution or any of the asset price
measurements, indicating that the distribution truly represents the steady state distribution of the economy.
5
Crisis Episode: Scenario 1
5.1
Risk premium and Sharpe ratio
Figure 1 graphs the risk premium and Sharpe ratio for the two hedge fund calibrations (m = 4, 6) as a
function of the scaled specialist-wealth (w/D). We plot these measures against w/D rather than a function
of the household’s scaled wealth, y = w h /D, in order to more clearly discuss the effects of the intermediation
constraint.13 w/D can be interpreted as the capital of the intermediation sector.
The prominent feature of our model, clearly illustrated by the graphs, is the asymmetric behavior of the
risk premium and Sharpe ratio. The right hand side of the graphs represent the unconstrained states of the
economy, while the left hand side represent the constrained states. The cutoff for the constrained region for
13 It
` ´
` ´
` ´
is easy to check that the boundary condition (20) F yb = yb implies that F 0 yb = 1, and if F yb > yb we have
` ´
F 0 yb > 1 (see Appendix A for details). Because in our numerical solutions we have F 0 (0) < 0, (20) ensures that the scaled
intermediation wealth w/D = F (y) − y is strictly decreasing in the scaled household’s wealth y. Thus we present our results
as functions of w/D to highlight the effect of specialist wealth.
26
Figure 1: Risk Premium and Sharpe Ratio
Risk Premium
Sharpe Ratio of the Stock
10
0.8
m=4
m=6
m=4
m=6
8
0.7
0.6
6
0.5
0.4
4
0.3
0.2
2
0.1
0
0
1
2
3
Scaled Intermediary Capital w/D
0
0
4
1
2
3
Scaled Intermediary Capital w/D
4
Risk premium (left panel) and Sharpe ratio (right panel) are graphed against scaled-specialist wealth (w/D). Parameters are m = 4, λ = 0.5 and m = 6, λ = 0.5 and those given in Table 2. The cutoff for the constrained region for the
two cases are 2.8 (m = 4 case) and 1.9 (m = 6) case.
the two cases are 2.8 (m = 4 case) and 1.9 (m = 6) case. Risk premia and Sharpe ratio rise as specialist
wealth falls in the constrained region, while being relatively constant in the unconstrained region.
This asymmetric behavior is intuitively what one would expect from the model: the model’s intermediation constraint is by its nature asymmetric, binding only when specialist wealth is low. To sharpen
understanding of the mapping between the constraint and risk premia, consider the following calculation.
As noted above, the pricing kernel in our model can be expressed in terms of the specialist’s consumption.
Thus, the risk premium on the risky asset is equal to:
γ covt
dct
, dRt
ct
To a first approximation, the volatility of the specialist’s consumption growth process is equal to the volatility
of his wealth return process (the approximation is exact if γ = 1). Thus,
vart
dct
ct
≈ αIt
2
vart (dRt ),
where αIt is the portfolio exposure to the risky asset in the intermediary’s (and specialist’s) portfolios.
Therefore, the risk premium is approximately,
γαIt vart (dRt) .
27
In our model, the variance of returns is roughly constant as a function of state (see the discussion of this
point below). Most of the action in the risk premium comes from the changing αIt . We have noted before
that in the constrained region, as households withdraw from intermediaries and limit their participation in
the risky asset market, the specialists increase their exposure to the risky asset (see equation (19)). This
dynamic, driven through αIt , explains the behavior of the risk premium. Figure 2 graphs αI as a function of
specialist wealth. We note the close correspondence between this graph and those in Figure 1.
Figure 2: Portfolio Holdings
Intermediary’s Position in the Risky Asset (αI)
100
m=4
m=6
80
60
40
20
0
0
1
2
3
Scaled Intermediary Capital w/D
4
The intermediary’s portfolio share in the risky asset (αI ) is graphed in the left panel against the scaled-intermediary
wealth (w/D). Parameters are m = 4, m = 6 and those given in Table 2.
Figures 1 and 2 are graphed for the two cases, m = 4 and m = 6. The cutoff for the constrained region
for these two cases are 2.8 (m = 4 case) and 1.9 (m = 6) case. Thus, as expected, the larger m leads to a
narrower constrained region. Comparing the slopes of the risk premium graph, in the constrained region,
for the m = 4 and m = 6 cases, the higher m case resembles a “convex” transformation of the lower m case.
In particular, deep into the constrained region, the risk premium is more sensitive to changes in specialist
wealth when m is larger.
Quantitatively, as one can see from Figure 1, the calibration produces a risk premium in the unconstrained
region of around 5%. The numbers for the risk premium are higher in the constrained region; however,
28
without knowing the probability that a given specialist-wealth state may occur, it is not possible to interpret
a statement about how much higher. To provide some sense for the values of the risk premium we may
be likely to observe in practice, we need to simulate the model and compute the equilibrium probability of
each state. We simulate the model as described in Section 4.4. The resulting steady state distribution over
specialist wealth is graphed in Figure 3 (for the m = 4 case). Also superimposed on the figure in dashed
lines is the risk premium from the previous graph.
Figure 3: Steady State Distribution
0
0.9
2
4
6
8
10
12
0.18
0.8
0.16
0.7
0.14
0.6
0.12
0.5
0.1
0.4
0.08
0.3
0.06
0.2
0.04
0.1
0.02
0
0
2
wc/D=2.80
4
6
8
10
0
12
Scaled Intermediary Capital w/D
The steady state distibution of w/D is graphed for the m = 4 case. The dashed line graphs the risk premium in order
to illustrate the actual range of variation of the risk premium.
Table 3 provides a number of statistics from this simulation. First note the leverage and income ratio
that we have used to pin down l and λ. The average income ratio from the data was between 30 − 35%;
our numbers in the table are closer to 37% and 40%. The leverage ratio suggested by the data was around
0.43. This number is close to the leverage ratio conditional on being in the unconstrained region. Clearly,
leverage rises in the constrained region.
The economy spends most of the time in the unconstrained region (59% and 68% for the two cases).
We may think of the unconstrained region as a “normal” non-crisis period. The average risk premium and
Sharpe ratio, conditional on being in the unconstrained region, is around 4.9% and 40% for each of the cases.
In the constrained region, the risk premium rises. The probability that the risk premium will exceed 7.5%
is around 4%. For the risk premium to exceed 10%, which is about double the unconstrained region average
29
Table 3: Measurements for Scenario 1
Panel A: Constrained and Unconstrained Regions
This panel presents a number of average measurements for the economy, broken down into conditional
on being in the constrained region, conditional on being in the unconstrained region, and unconditional
average. Parameters for the two cases reported are given in Table 2.
m = 4 Case
m = 6 Case
Avg. Unconst. Const.
Avg. Unconst. Const.
Probability
59.09
40.91
67.75
32.25
Risk Premium (%)
5.34
4.87
6.01
5.32
4.97
6.03
Sharpe Ratio (%)
42.91
39.22
48.24
43.06
40.34
48.97
Interest Rate (%)
0.50
0.88 −0.045
0.50
0.79 −0.099
Leverage Ratio (%)
47.56
41.43
56.42
47.67
42.86
57.72
Income Ratio (%)
36.89
40.58
31.55
40.07
41.76
36.36
Panel B: Measures at Different Risk Premia
The different risk premia at which we compute the various measures are given in the first row (denoted π).
The second row reports the probability that the economy will ever reach a value of risk premium greater
than the given π. The rest of the rows report measures at the given π.
m=4
m=6
Risk Premium (%) ≡ π
5%
7.5%
10%
20%
5%
7.5%
10%
20%
Prob (Risk Premium> π) 34.37
4.35
1.04
0.05 63.31
3.33
0.81
0.04
Sharpe Ratio at π
40.35
59.93
80.60 170.82 40.75
60.76
81.21
169.63
Interest Rate at π
0.84
−1.36
−3.71 −13.82
0.84
−1.46
−3.81
−13.92
Leverage Ratio at π
47.55
70.02
79.92
92.24 48.56
70.94
80.33
92.10
in terms of both risk premium and Sharpe ratio, the probability is 1%. An extreme crisis that increases risk
premia and Sharpe ratio about 4.5X to 20% is very unlikely. Our model puts this probability around 0.05%.
Table 3 also provides a sense as to the effect of varying m, by comparing the two cases represented.
There are two, almost offsetting effects of m. Raising m lowers the probability of the constrained region
from 41% to 32%. However it increases the probability that the risk premium will exceed 5% from 34% to
63%, while leaving the probabilities at the more extreme points relatively unaffected. The constrained region
gets smaller, but the probability mass becomes concentrated at a slighly larger value of the risk premium.
The net effect is almost a wash as the average risk premium across the two cases is within two basis points.
To put these numbers in perspective, consider the 1998 crisis. Figure 4 graphs the behavior of the high
grade credit spread (AAA bonds minus Treasuries), the spread on FNMA mortgage backed securities relative
to Treasuries, and the option adusted spread on volatile interest-only mortgage derivative securities (data
are from Gabaix, Krishnamurthy, and Vigneron, 2007). The spreads are graphed over a period from 1997 to
1999 and includes the fall of 1998 hedge fund crisis. During 1997 and upto the middle of 1998 spreads move
in a fairly narrow range. If we interpret the unconstrained states of our model as this “normal” period, then
the muted response of risk premia to the state can capture this pre-crisis period. In a short period around
30
Figure 4: Crisis Spreads
2500
200
180
160
1500
1000
500
140
Yield Spread (bps)
OAS Spread (bps)
2000
120
100
80
M
ar
Ap 97
r
M -97
ay
-9
Ju 7
n9
Ju 7
l
Au -97
gSe 97
pO 97
ct
N -97
ov
D -9 7
ec
Ja 97
n
Fe -98
bM 98
ar
Ap 98
r
M -98
ay
Ju 98
n9
Ju 8
l
Au -98
gSe 98
pO 98
ct
N -98
ov
D -9 8
ec
Ja 98
n
Fe -99
bM 99
ar
Ap 99
r
M -99
ay
-9
Ju 9
n9
Ju 9
l
Au -99
gSe 99
pO 99
ct
-9
9
0
IO OAS
CREDIT
MBS
The spreads between the Moody’s index of AAA corporate bonds and the 10 year Treasury rate (grey line), the
spreads between FNMA 6% TBA mortgage-backed securities and the 10 year Treasury rate (black line), and the
option-adjusted spreads on a portfolio of interest-only mortgage-backed securities relative to Treasury bonds (dashed
line) are graphed monthly from 1997 to 1999.
October 1998 spreads on these securities increase sharply. The credit spreads and MBS spreads double from
their pre-crisis level. There is manifold increase on the mortgage derivative spread. Although it is hard to
estimate precisely how much Sharpe ratios increase during the episode, a doubling is plausibly within the
range of estimates. Certainly from the standpoint of standard representative household models, even a 50%
increase during the 1998 event is difficult to understand as aggregate consumption was barely at risk. In
our model, the asymmetry in the intermediation constraint calibrated to hedge fund data can generate the
dramatic increase in risk premia around crises. Many observers comment on the reduction in intermediation
capital in the fall of 1998 crisis, and refer to the episode as a tail event. Both statements make sense from
our calibration.
5.2
Flight to quality
Figure 5 graphs the interest rate as a function of specialist wealth. As in previous graphs, there is an asymmetric effect evident: a relatively constant interest rate in the unconstrained region, and a falling interest rate
in the constrained region. There are two intuitions behind the interest rate effect in the constrained region.
31
Figure 5: Interest Rate
Interest Rate
0.5
m=4
m=6
0
−0.5
−1
0
1
2
3
Scaled Intermediary Capital w/D
4
Interest rate is graphed against scaled-specialist wealth (w/D). Parameters are m = 4 and m = 6 and those given in
Table 2.
First, as the specialist’s consumption volatility rises with the tightness of the intermediation constraint, the
precautionary savings effect increases specialist demand for the riskless bond. Second, as specialist wealth
falls, households withdraw equity from intermediaries, increasing their demand for the riskless bond. To
clear the bond market, the equilibrium interest rate has to fall.
Both the behavior of the interest rate and the disintermediation-driven demand for bonds is consistent
with a flight to quality. Since the total funds under intermediation is equal to (1 + m)wt in the constrained
region, a fall in wt leads to m times larger fall in household funds invested in intermediaries. Moreoever, since
the intermediary holds a levered position in the risky asset, the sensitivity of fund flows to the risky asset
price increases with the tightness of the intermediation constraint. Formally, we can compute the percentage
outflow of household funds,
dwth
(1+m)wt ,
as a function of the percentage change in risky asset price as:
dwth
m dP I
=
α .
(1 + m)wt
1+m P
(21)
We note that the elasticity is proportional to αI which is increasing in the tightness of constraints.
The row in Table 3 corresponding to the interest rate provides a quantitative sense for the variation in
the interest rate from the simulations. The unconditional average interest is 0.50% which is consistent with
empirical estimates. However, the interest rate is over-sensitive to the state in our model. At the 7.5% risk
32
premium state, the interest rate is around −1.40%, falling to −3.7% at the 10% risk premium state (for the
m = 4 case). The intuition offered in the previous paragraphs helps to explain why. Both agents increase
their demand for bonds in the constrained region; but since the bond is in zero net supply, the interest rate
has to adjust dramatically. In practice, the bond that investors fly-to during a crisis episode (i.e. government
bonds) is not in zero net supply, which may be one reason why the interest rate effect in our model is at
odds with the data. Later in this paper we will explore an extension of the model with government bonds
in positive supply and the interest rate effects will be dampened.
5.3
Volatility
Figure 6: P/D Ratio and Volatility
Price/Dividend Ratio F(.)
Volatility of Stock Price
27.5
27
0.14
m=4
m=6
0.13
0.12
26.5
m=4
m=6
0.11
26
0.1
25.5
0.09
25
0.08
24.5
24
0
0.07
2
4
6
8
Scaled Intermediary Capital w/D
0.06
0
10
5
10
15
20
25
30
Scaled Intermediary Capital w/D
35
Price/Dividend ratio (left panel) and stock return volatility (right panel) are graphed against scaled specialist wealth
(w/D). Parameters are m = 4 and m = 6 and those given in Table 2.
The left-hand panel of Figure 6 graphs the price/dividend ratio as a function of specialist wealth. Consistent with intuition, over most of the range, F (·) falls as specialist wealth falls. However, there is a
non-monotonicity that arises when w/D is very small.
This effect arises because interest rates diverge to negative infinity when w/D approaches zero. There
are two forces affecting the discount rates applied to dividends in determing F (·): On the one hand, the risk
premium is high when w/D is small; on the other hand, the interest rate is low when w/D is high. These
two effects combine to produce the non-monotonicity of F (·) when w/D is small.
33
The right-hand panel of Figure 6 graphs the stock return volatility as a function of specialist wealth.
The volatility is close to constant and actually falls in the constrained region. The latter effect is driven by
the non-monotonicity in F (·). The stock price is equal to Dt × F (yt ). The non-monotonicity means that a
shock that causes a fall in Dt leads to a rise in F (yt is negatively correlated with Dt ). This is a failure of
the model: we are unable to replicate the observed increase in conditional volatility accompanying a crisis
period.
It is worth pointing out that despite the fall in conditional volatility in the crisis region, the risk premium
rises. This highlights that the risk premium is driven largely by the endogenous increase in the volatility of
the pricing kernel (γσ( dc
)) in our model.
c
5.4
Capital movement and recovery from crisis
Referring to Figure 4, the corporate bond spread and MBS spread widen from 90 bps in July 1998 to a high of
180 bps in October 1998 before coming down to 130 bps in June 1999. Thus, the half-life — that is, the time
it takes the spread to fall halfway to the pre-crisis level — is about 10 months. The interest-only mortgage
derivative spread, which is very sensitive to market conditions, widens from 250 bps in July 1998 to a high
of 2000 bps before coming back to 500 bps in June 1999. We note that this timescale for mean reversion,
on the order of months, is much slower than the daily mean-reversion patterns commonly addressed in the
market micro-structure literature (e.g., Campbell, Grossman, and Wang, 1994).
A common wisdom among many observers is that this pattern of recovery reflects the slow movement of
capital into the affected markets (Froot and O’Connell, 1999, Berndt, et. al., 2004, Mitchell, Pedersen, and
Pulvino, 2007, Duffie, Garleanu, and Pedersen, 2007). Our model captures this slow movement. We will
show in this section that our baseline calibration can also replicate these speeds of capital movement.
In the crisis states of our model, risk premia are high and the specialists hold leveraged positions on
the risky asset. Over time, profits from this position increase wt , thereby increasing the capital base of
the intermediaries. The increase in specialist capital is mirrored by an m-fold increase in the allocation of
households’ capital to the intermediaries, as the intermediation constraint is relaxed. Together these forces
reflect a movement of capital back into the risky asset market and lead to increased risk-bearing capacity
and lower risk premia. Note, however, that one dimension of capital movement that plausibly occurs in
practice but is not captured by our model is the entry of “new” specialists into the risky asset market.
We can use the model simulation to gauge the length and severity of a crisis within our model. Table 4
34
presents data on how long it takes to recover from a crisis in our model. We fix a state (y, D) corresponding to
an instantaneous risk premium in the “Transit from” row. Simulating the model from that initial condition,
we compute and report the first passage time that the state hits the risk premium corresponding to the
“Transit to” column. The time is reported in years.
If we start from the extreme crisis state of 20% and
Table 4: Crisis Recovery
This table presents transition time data from simulating the model for the m = 4
and m = 6 cases. The rest of the parameters are those given in Table 2. We fix
a state corresponding to an instantaneous risk premium of 20% (“Transit from”).
Simulating the model from that initial condition, we compute and report the first
passage time that the state hits the risk premium corresponding to that in the
“Transit to” column. Time is reported in years. The column “Increment time”
reports the time between incremental “Transit to” rows.
m=4
m=6
Transit from 20 Increment time Transit from 20 Increment time
Transit to
15
0.23
0.23
0.21
0.21
12.5
0.46
0.23
0.43
0.22
10
1.02
0.56
0.93
0.5
7.5
2.62
1.6
2.48
1.55
6
5.91
3.29
5.46
2.98
5
12.88
7.1
12.56
6.97
compute how long it takes to recover to 12.5% — i.e. halfway back to the unconditional average levels we
report earlier of around 5% — the time is 0.46 years (5.5 months) for the m = 4 case. The transit times are
uniformly faster for the m = 6 case because as specialist capital increases, households react more strongly
in bringing their own capital back to intermediaries.
From the 10% crisis state to the 7.5% state takes 1.6 years in the m = 4 case. For the fall of 1998 episode,
the half-life we suggested was around 10 months. The model half-life from 10% is larger, but is in the order
of magnitude of the empirical observation. It is worth keeping in mind that it is difficult to measure exactly
how high risk premia or Sharpe ratios were at the peak of the crisis. For example, if we judge that the peak
risk premia corresponded to 13% in our model, then the half-life down to 8% is close to 10 months (not
reported in the table).
One failing of the model that we see from Table 4 is the extremely slow recovery from 6% to 5%. Our
simulations put these numbers around 7 years. Essentially, the specialist profits when the risk premium is 6%
are so small that specialist capital does not grow very fast and hence our model predicts a correspondingly
slow recovery time. In practice, the final stages of recovery from crisis may also lead to other “specialists”
moving into the affected market. Our model does not capture such an effect. It is also worth noting that
35
in practice, statistically distingushing a 6% risk premium from a 5% risk premium is difficult, so that the
recovery time from 6% to 5% may not be a meaningful measurement.
The slow adjustment of risk premia, in timescales of many months, during the 1998 episode is also
consistent with other studies of crisis episodes. Berndt, et. al. (2005) study the credit default swap market
from 2000 to 2004 and note a dramatic market-wide increase in risk premia (roughly a quadrupling) in
July 2002 (see Figures 1 and 2 of the paper). Risk premia gradually fall over the next two years: From
the peak in July 2002, risk premia halve by April 2003 (9 months). The authors argue that dislocations
beginning with the Enron crisis led to a decrease in risk-bearing capacity among corporate bond investors.
Mirroring the decreasing risk-bearing capacity, risk premia rose before slowly falling as capital moved back
into the corporate bond market and expanded risk bearing capacity. Gabaix, Krishnamurthy, and Vigneron
(2007) note a dislocation in the mortgage-backed securities in late 1993 triggered by an unexpected wave of
consumer prepayments. A number of important hedge fund players suffered losses and went out of business
during this period, leading to a reduction in risk bearing capacity. Figure 3 in the paper documents that
risk premia reached a peak in December 1993 before halving by April 1994 (5 months). Froot and O’Connell
(1999) study the catastrophe insurance market and demonstrate similar phenomena. When insurers suffer
losses that deplete capital they raise the price of catastrophe insurance. Prices then gradually fall back to
long-run levels as capital moves back into the catastrophe insurance market. Froot and O’Connell show that
the half-life in terms of prices can be well over a year.14
Each of these markets are intermediated markets that fit our model well. Investors are institutions who
have specialized expertise in assessing risk in their markets. Our theory explains the slow movement of risk
bearing capacity and risk premia documented in these case-studies. The calibrated model also captures the
frequency of the slow adjustment of risk premia.
6
Aggregate Asset Pricing: Scenario 2
The second exercise we perform with our model is an aggregate asset pricing calibration. Our model is
closely related to heterogeneous agent models that have been proposed in the literature to explain the equity
premium puzzle (for example, see Mankiw and Zeldes (1991) and Vissing-Jorgensen (2002)). In these models,
14 Mitchell,
Pedersen, and Pulvino (2007) document similar effects in the convertible bond market in 1998 and again in 2005.
In both cases, crisis recovery times are in the order of months. They also note that spreads in merger arbitrage strategies took
several months to recover following the October 1987 stock-market crash.
36
as in ours, the marginal investor’s consumption is more volatile than that of the average investor. Thus,
these models are able to generate a pricing kernel that is more volatile than that of the aggregate endowment,
which is the useful property in explaining the equity premium puzzle.
We suppose that the risky asset of the model encompasses all risky financial assets including stocks,
mortgages, etc., and that all investment in these assets is made by intermediaries. While clearly in practice
there are investments in risky assets that do not require the expertise of an intermediary, the exercise provides
a benchmark for our intermediation model. Allen (2001) notes that over 50% of financial assets are held
through intermediaries. He also observes that many investment decisions over the remaining 50% of financial
wealth are made with the advice of professional advisors. Plausibly, investors will place less weight on the
advice of a specialist whose own wt has been falling. Then, if wt is falling, households withdraw from the
risky market, so that risk is concentrated among the specialists – reflecting the main dynamic of our model.
6.1
m = 1 case
The left-panel of Figure 7 graphs the steady state distribution of w/D for the m = 1 case. Contrasting
this figure with Figure 3 we see that most of the probability mass is in the left side of the graph, which
corresponds to the constrained region, but the probability mass in the extreme crisis region is small. The
probability mass in the constrained region is 98% for this case, suggesting that there is almost always some
feedback from intermediary capital to household participation. The right-panel of Figure 7 graphs the risk
premium for the m = 1 case. We also graph the risk premium for the earlier m = 4 case as constrast. The
risk premium for the m = 1 case rises more gradually as w/D falls, capturing a more muted response to the
state.
6.2
Asset Price Measurements
Table 5 presents the results from simulating the model and computing a number of unconditional moments
for the m = 1 case, as well as a variation with m = 0.5. Let us focus on the m = 1 case first. Our choice of
λ produces a leverage ratio of 0.46. The data on which we based our choices of λ suggested a leverage ratio
in this range. Our choice of l = 1 produces the financial to total income ratio for the household of 0.32. The
data we cited earlier suggests numbers between 0.30 and 0.35.
The average risk premium for the m = 1 case is 5.63%. There are many points of comparison for this
risk premium. First, this number is within the range of estimates of the equity risk premium, which is one
37
Figure 7: Steady State Distribution and Risk Premium
Risk Premium
0.1
0.9
m=1
0.09
0.8
0.08
m=4
m=1
0.7
0.07
0.6
0.06
0.5
0.05
0.4
0.04
0.3
0.03
0.2
0.02
0.1
0.01
0
0
3
6
9 wc/D=9.93 12
Scaled Intermediary Capital w/D
15
18
0
0
2
4
6
8
Scaled Intermediary Capital w/D
10
12
The left panel graphs the steady state distibution of w/D for the m = 1, λ = 0.4 case. The right panel graphs the
risk premium for both m = 1, λ = 0.4 and m = 4, λ = 0.5 cases.
Table 5: Measurements for Scenario 2
m=1
λ = 0.4
Risk Premium (%)
5.63
std[Risk Premium] (%)
12.89
Interest Rate (%)
0.29
std[Interest Rate] (%)
1.14
Sharpe Ratio (%)
44
Intermediary Leverage (%)
46
Household Financial Income/Total Income (%)
32
Prob of Constraints (%)
97.7
m = 0.5
λ = 0.4
5.91
13.46
0.05
1.02
43.40
45
22.59
99.91
guide as to the expected return on aggregate wealth. Second, from a model standpoint if the risky asset
of our model was priced by a representative agent who consumes NIPA aggregate consumption and has a
coefficient of relative risk aversion of two, then the highest risk premium that such an agent would require to
hold the risky asset would be 2 ×0.12 ×0.038. That is, if the risky asset is perfectly correlated with aggregate
consumption, and using consumption growth volatility of 3.8%, we would compute a risk premium of 0.91%.
It is likely that the payoff on the typical intermediated asset is less than perfectly correlated with NIPA
consumption growth (Campbell and Cochrane (1999) estimate a 0.15 correlation of stock market dividends
with aggregate consumption), so in the representative household model the risk premium is even smaller.
Moving from a representative household analysis to one which recognizes that intermediaries are marginal
38
in pricing intermediated assets immediately means that the pricing kernel is perfectly correlated with intermediated payoffs and has volatility corresponding to those of intermediated payoffs. That is, consider a
standard Lucas-tree model with dividend volatility of 12% and a representative agent with risk aversion of
two. The risk premium in such a model is 2 × 0.12 × 0.12 = 2.88%.
Our model produces a higher risk premium than each of these benchmarks. It does so by additionally
incorporating leverage and constraints in intermediation that lead intermediaries to increase leverage and
have a more concentrated exposure to the returns on intermediated payoffs, when constraints bind.
The volatility of the excess return on the risky asset is 12.89%. As a benchmark, note that dividends
are calibrated to 12% volatility, so that in a standard model we would expect to see volatility of close to
12%. Our model produces more volatility than the 12% benchmark, but not much more. Since the model
yet produces a high risk premium, it is clear that the model works through a volatile pricing kernel. The
Sharpe ratio in the model is around 44%.
Finally, the model’s riskless interest rate averages 0.29% is close to that observed empirically. The riskless
rate volatility ranges is 1.14% which is similar to numbers reported in Campbell (1999).
Table 5 presents results for the m = 1 case as well as a variation with m = 0.5 case to give a sense as
to the sensitivity of the simulation to varying m. The risk premium and Sharpe ratios are higher for the
lower m case. Note that the probability of the constrained region rises to 99.91 percent for the m = 0.5
case. Lowering m tightens the intermediation constraint and reduces participation for every level of specialist
wealth. Thus, lowering m, on average raises the risk premium. Note the contrast between this effect and
one we find in the hedge fund scenario: we show there that, conditional on being constrained, the economy
recovers faster if m is larger.
6.3
Liquidity and time varying expected returns
There is a great deal of evidence in finance for time variation in investors’ required returns on risky assets.
In the model, despite dividends being i.i.d., there is time variation in risk premia. This is because specialist
wealth varies over time and the risk premium is a function of specialist wealth. The following computation
provides a quantitative sense of the time variation in risk premia generated by the model. If we take the
mean state in the simulation, and move one standard deviation towards a more constrained state the risk
premium rises to 6.38%. Moving one standard deviation towards a less constrained state causes the risk
premium to fall to 4.32%. Thus, 2.06% variation in total.
39
In principle, if we can empirically measure “specialist wealth,” these measurements can be used to predict
asset returns. Mapping our model to practice, specialist wealth most naturally corresponds to the financial
health of the intermediation sector, so that the conditioning variable should be constructed from the balance
sheets of the intermediation sector. Unfortunately, we are unaware of any empirical research that has
specifically pursued this link.
The closest research in this vein are papers that document a role for a liquidity factor in explaining
asset returns. Pastor and Stambaugh (2003) and Sadka (2006) report high Sharpe ratios of around 0.75
from a strategy of going long “low liquidity” stocks and short “high liquidity” stocks (see also Amihud,
2002, and Acharya and Pedersen, 2003). Low (high) liquidity stocks are measured to be stocks whose
returns covary positively (negatively) with marketwide illiquidity conditions. Their results have received
considerable attention in the literature and raise the possibility that liquidity may be a priced factor. The
factor that these papers construct is based on a micro-structure measure of price impact. One natural
interpretation of this measure is in terms of the capacity of market makers and other financial market
specialists to provide liquidity to markets. In particular, if financial market specialists have the capacity to
absorb non-informational trade then price impact will be low. However, if financial market specialists are
capital constrained or extremely risk averse then price impact will be high.
Our model applies to intermediaries pricing intermediated payoffs. Although our model does not have a
micro-structure element, providing liquidity to markets is the principal function of intermediaries. Thus we
may interpret our model to say that when intermediation constraints are tight, markets will be illiquid in
the sense of Pastor and Stambaugh (2003) and Sadka (2006). That is, the micro-structure based liquidity
measures of Pastor and Stambaugh (2003) and Sadka (2006) should reflect variation in wt .
Our results thus offer one explanation for the findings in these papers. In turn, this research also suggests
that one way to measure intermediation constraints may be to look to asset market liquidity measures. In
the next section of the paper, we more explicitly draw a connection between our model and liquidity.
7
Debt and Market Liquidity
The risky asset in the model is illiquid in the sense that only the specialists participate in the risky asset
market. The riskless asset is liquid since all agents participate in the market. From this standpoint, the
risk premium of our model is at least partly a liquidity premium. This notion of liquidity, deriving from
40
market segmentation, is similar to Allen and Gale (1994) or models of the liquidity of money in the monetary
economics literature (e.g., Alvarez, Atkeson, and Kehoe, 2002). So far we have shown how changing intermediary capital endogenously affects participation and the liquidity premium on the risky asset. Particularly in
the calibration of scenario 1, we have shown that when w falls low enough the model can replicate a liquidity
crisis event.
In this section, we consider a different experiment that alters the liquidity of the risky asset. We allow
for the riskless asset to be in positive supply, and ask how changing the supply of the riskless asset affects
the premium on the risky asset. Since the riskless asset is the liquid asset, this experiment is akin to “adding
liquidity” to the market.
This experiment is useful for two reasons. First, by demonstrating that altering the supply of the riskless
asset affects the risk premium on the risky asset, we clarify that the risk premium on the risky asset is
in part a liquidity premium. Krishnamurthy and Vissing-Jorgensen (2007) present empirical evidence that
changes in the supply of government debt have a causal effect on the spread between corporate bond yields
and Treasury bond yields. They interpret their results as showing that part of the corporate-Treasury
spread is a liquidity premium. Our theoretical results help to shed light on their empirical findings. Second,
the crisis-intermediation literature (e.g., Holmstrom and Tirole, 1998, or Diamond and Rajan, 2005) and
the monetary economics literature (e.g., Alvarez, Atkeson, and Kehoe, 2002) commonly study how adding
“liquidity” to the economy alters asset prices and allocations. In the existing literature, adding liquidity
lowers the liquidity premia on less liquid assets. As we show, our model can replicate this effect.
Formally, we introduce a government into the model that rolls over a short-term riskless government
bond, whose interest cost is financed partly by levying lumpsum taxes on the households. Since the model
assumes an overlapping generation structure for households, the model is not Ricardian, and this type of
government bond policy can affect the equilibrium.
We assume that at date t the government has bonds outstanding of B(yt , Dt ) on which it pays the interest
rate of rt . Then, the flow budget constraint for the government is,
dBt − rt Bt dt + τt dt = 0.
In this equation, dBt is the net increase in bond issuance (i.e., moneys received from issuance in excess of
moneys paid to redeem the existing issue), rt Bt is the interest cost on the outstanding stock of debt, and τt is
the lumpsum taxes raised on households. The tax affects the wealth dynamics for the households. Equation
41
(5) is altered to,
f t − rt dt − τt dt.
dwth = (lDt − ρwth )dt + wth rt dt + αht (1 − λ)wth dR
We consider the following class of bond policies that are mathematically easy to analyze in our setting,
B(yt , Dt ) = Dt max[0, b(yt − y∗ )]
for constants, b and y∗ . The policy scales the bond issue linearly with dividends. Scaling with respect to an
endogenous variable is obviously important in order that the bond policy does not either dominate or vanish
from the economy. In addition, since the key dynamic of the model occurs in the constrained region when yt
is high (and wt is low), it is interesting to study a policy whereby more bonds are issued when the economy
is more constrained. Thus we additionally scale the bond issue with the variable yt − y∗ .
With these policies, the ODE of Proposition 1 remains the same. The coefficents on the dynamics of
scaled household wealth, µy and σy , are altered due to taxes. We derive the new µy and σy in the Appendix.
The boundary condition for the economy also changes. Previously, at the upper boundary for y we have:
yb = F (yb ).
That is, the household owns all of the wealth of the economy, amounting to F (y). With the introduction of
government bonds, this boundary changes to:
yb = F (yb ) + B(yb )/D.
(22)
We note that increasing B(yb ) causes F (yb ) to fall. Loosely speaking, we may think of this as a “crowding
out” effect.
In Table 6, we present results for different values of y∗ and b, and for the m = 4 baseline case. To normalize
across these different cases, we also report the average simulated value of government bonds divided by total
wealth (risky asset plus government bonds). As a benchmark, currently in the US, total wealth is around
$50tn and total government debt is $5tn, giving a ratio of 10%.
Comparing across the first two rows of the table we see that adding debt raises interest rates and lowers
the risk premium. The interest rate effect should be expected since the model is not Ricardian. We also
note that having the debt security in positive supply softens the fall in the interest rate in the constrained
region. The constrained average interest rate is 0.77% rather than −.045% in the b = 0 case.
The risk premium falls on average by 12 bps. The effect is larger if we condition on a given level of
the constraint. For example, the constrained average risk premium falls by 34 bps, while the unconstrained
42
Table 6: Debt and Liquidity
This panel presents a number of average measurements illustrating the effect of introducing government
debt into the economy. The measures are broken down into conditional on being in the constrained region,
conditional on being in the unconstrained region, and unconditional average. Parameters are the m = 4
case and those given in Table 2. Debt parameters are given in the first column of the table
Risk Premium (%)
Interest Rate (%)
h
i
B
Avg. Unconst. Const. Avg. Unconst. Const. Prob(Const.)
E P +B
b=0
5.34
4.87
6.01 0.50
0.88 −0.045
0.41
0
b = 0.20, y∗ = 10 5.22
4.61
5.67 1.10
1.56
0.77
0.58
0.10
b = 0.30, y∗ = 10 5.14
4.45
5.50 1.49
1.94
1.25
0.65
0.176
b = 0.15, y∗ = 5
5.19
4.56
5.64 1.12
1.61
0.78
0.58
0.10
b = 0.35, y∗ = 15 5.32
4.82
5.85 1.11
1.32
0.87
0.49
0.10
average premium falls by 26 bps. Figure 8 graphs the risk premium and interest rate as a function of scaled
specialist wealth for the no debt case (b = 0) and the debt case corresponding to the second row of Table
6. The figure shows that the risk premium is uniformly lower when debt is in the model. Although not
apparent from the figure, the difference in the two risk premia curves is about 50 bps around the point w c /D.
To understand why the average risk premium falls less than 50 bps, note the differences in the probability
of the constrained region, given in the penultimate column of the table. This probability rises from 0.41
to 0.58. Adding debt shifts the entire steady state distribution of the economy more into the constrained
region. This occurs because government debt raises the interest rate, and since the specialist is on average
borrowing in the debt market, the higher interest rate lowers his wealth on average. Hence, the economy is
more likely to fall into the constrained region. These two effects of debt work against each other, but the
net effect reduces the average risk premium by 12 bps.
Row 3 of Table 6 presents the results for the case of a 17.6% debt ratio. The risk premium falls by 8 bps
relative to the 10% debt case, which suggests that the marginal effect of debt on the risk premium falls as
the level of debt rises.
The last two rows of the Table consider variations on y∗ . We consider values of y∗ of 5 and 15, adjusting
b so that the average level of debt in the economy remains at 10%. The case of y∗ = 5 is almost identical to
the case of y∗ = 10. The economy spends little time at these values of y so that altering debt in this region
has limited effects on the equilibrium. The case of y∗ = 15 looks closest to the b = 0 case, with a small risk
premium effect. This result can be understood be returning to the boundary condition (22). Since the value
of b is highest for the y∗ = 15 case, and our debt policy is equal to b(yb − y∗ ), the y∗ = 15 case also leads
to the highest value of B(yb , D). From (22) we see that a large value of B(·) lowers the boundary value of
43
Figure 8: Debt, Interest Rate, and Risk Premium
Interest Rate
Risk Premium
0.02
0.2
0.18
0
0.16
b=0
b=0.2035,y*=10
−0.02
0.14
−0.04
0.12
b=0
b=0.2035,y*=10
−0.06
0.1
0.08
−0.08
0.06
−0.1
0
1
2
3
I
Scaled Intermediary Capital w /D
0.04
0
4
1
2
3
I
Scaled Intermediary Capital w /D
4
Interest rate (left panel) and risk premium (right panel) are graphed against scaled-specialist wealth (w/D). Two
curves are plotted in panel, corresponding to the cases of no government debt and a debt/wealth ratio of 10%. The
rest of the parameters are m = 4 and those given in Table 2.
F (yb ). In turn, the lower boundary value causes F (y) to fall faster for every value of y, leading to higher
price and return volatility. This factor tends to drive the risk premium up, and the net effect is that the risk
premium does not fall as much relative to the b = 0 case.
Taken together, these results show that adding liquid debt to the economy lowers the risk premium on
the risky asset. This result is consistent with other theoretical and empirical work and is therefore a success
of the model. The current debt/wealth ratio in the US is approximately 10%. Across the various rows of
Table 6 we see that adding debt to match a 10% debt/wealth ratio lowers the risk premium by between
10 and 15 bps. This suggests that the liquidity premium produced by the calibrated model is between 10
and 15 bps. A failure of the model is that, quantitatively, this number is too small. Krishnamurthy and
Vissing-Jorgensen (2007) suggest a liquidity premium of closer to 70 bps. Clearly more study is required to
bring the empirical and theoretical measurements closer together.
8
Conclusion
TO BE ADDED
44
References
[1] Acharya, Viral and Lasse Pedersen, 2003, Asset Pricing with Liquidity Risk, Journal of Financial
Economics, forthcoming.
[2] Allen, Franklin, 2001, Do Financial Institutions Matter?, Journal of Finance, 2001, 56, 1165-1175.
[3] Allen, Franklin and Douglas Gale, 1994, Limited Market Participation and Volatility of Asset Prices,
American Economic Review 84, 933-955.
[4] Allen, Franklin and Douglas Gale, 2005, “From Cash-in-the-Market Pricing to Financial Fragility,”
Journal of the European Economic Association, 3(2-3), 535-546.
[5] Allen, Franklin and Gary Gorton, 1993, Churning Bubbles, Review of Economic Studies, 60, 813-836.
[6] Alvarez, Fernando, Andrew Atkeson, and Patrick Kehoe, 2002, “Money, interest rates, and exchange
rates with endogenously segmented markets,” Journal of Political Economy 110, 73—112.
[7] Amihud, Yacov, 2002, Illiquidity and Stock Returns: Cross-Section and Time-Series Effects, Journal of
Financial Markets, 5, 31-56.
[8] Barberis, N., M. Huang, and T. Santos, 2001, Prospect theory and asset price, The Quarterly Journal
of Economics, 116, 1-53.
[9] Basak, Suleyman and Domenico Cuoco, 1998, An Equilibrium Model with Restricted Stock Market
Participation, Review of Financial Studies, 11, 309-341.
[10] Bates, David, 2003, Empirical Option Pricing: A Retrospection, Journal of Econometrics, 116, 387–404.
[11] Berndt, Antje, Rohan Douglas, Darrell Duffie, Mark Ferguson, and David Schranz, 2004, Measuring
Default Risk Premia from Default Swap Rates and EDFs, mimeo, Stanford University.
[12] Borodin, A.N. and P. Salminen, 1996, Handbook of Brownian Motion. Birkhauser, Boston, MA.
[13] Boudoukh, Jacob, Matthew Richardson, Richard Stanton, and Robert Whitelaw, 1997, Pricing
mortgage-backed securities in a multifactor interest rate environment: A multivariate density estimation
approach, Review of Financial Studies 10, 405-446.
[14] Brunnermeier, Markus and Lasse Pedersen, 2005, Market Liquidity and Funding Liquidity, working
paper, Princeton.
45
[15] Campbell, John, 1999, “Asset Prices, Consumption, and the Business Cycle”, Chapter 19 in John Taylor
and Michael Woodford eds., Handbook of Macroeconomics, Amsterdam: North-Holland.
[16] Campbell, John, Sanford Grossman and Jiang Wang, 1993, “Trading Volume and Serial Correlation in
Stock Returns,” Quarterly Journal of Economics, 108, 905-940.
[17] Chan, Nicholas, Mila Getmansky, Shane Haas, and Andrew Lo, 2005, Systemic Risk and Hedge Funds,
with Nicholas Chan, to appear in M. Carey and R. Stulz, eds., The Risks of Financial Institutions and
the Financial Sector. Chicago, IL: University of Chicago Press.
[18] Collin-Dufresne, Pierre, Robert Goldstein, and J. Spencer Martin, 2001, The Determinants of Credit
Spread Changes, The Journal of Finance, 66, 2177-2207.
[19] Dasgupta, Amil, Andrea Prat, and Michela Verardo, 2005, The Price of Conformism, working paper,
LSE.
[20] Diamond, Douglas, 1997, “Liquidity, Banks, and Markets,” Journal of Political Economy, 105(5), pp.
928-56.
[21] Diamond, Douglas and Philip Dybvig, 1983, “Bank Runs, Deposit Insurance, and Liquidity,” Journal
of Political Economy, 91(5), pp. 401-19.
[22] Diamond, Douglas and Raghuram Rajan, 2005, “Liquidity Shortages and Banking Crises,” Journal of
Finance 60 (2), pp. 615-647.
[23] Duffie, Darrell, Nicolae Garleanu and Lasse Pedersen, 2007, “Valuation in Over-the-Counter Markets,”
Review of Financial Studies, forthcoming.
0
[24] Froot, Kenneth A., and Paul G. O Connell, 1999, The Pricing of US Catastrophe Reinsurance in The
Financing of Catastrophe Risk, edited by Kenneth Froot, University of Chicago Press.
[25] Fung, William, and David Hsieh, 2006, “Hedge Funds: An Industry in Its Adolescence,” Federal Reserve
Bank of Atlanta Economic Review, (Fourth Quarter), 91, 1-33
[26] Gabaix, Xavier, Arvind Krishnamurthy and Olivier Vigneron, 2007, Limits of Arbitrage: Theory and
Evidence from the MBS market, Journal of Finance, 62(2), 557-596.
46
[27] Garleanu, Nicolae, Lasse Heje Pedersen, and Allen M. Poteshman, 2005, Demand-Based Option Pricing,
Wharton mimeo.
[28] Gromb, Denis and Dimitri Vayanos, 2002, Equilibrium and Welfare in Markets with Financially Constrained Arbitrageurs, Journal of Financial Economics 66, 361-407.
[29] Grossman, Sanford J. and Zongquan Zhou, 1996, Equilibrium Analysis of Portfolio Insurance, Journal
of Finance 51, 1379-1403.
[30] He, Zhiguo, and Arvind Krishnamurthy, 2006, “Intermediation, Capital Immobility, and Asset Prices,”
working paper, Northwestern University.
[31] Holmstrom, Bengt and Jean Tirole, 1997, Financial Intermediation, Loanable Funds, and the Real
Sector, Quarterly Journal of Economics 112, 663-691.
[32] Holmstrom, Bengt and Jean Tirole, 1998, “Private and Public Supply of Liquidity,” Journal of Political
Economy, 106(1), 1-40.
[33] Karatzas, I. and S. Shreve, 1998, Methods of Mathematical Finance. New York: Spring-Verlag.
[34] Krishnamurthy, Arvind, 2002, “The Bond/Old-Bond Spread,” Journal of Financial Economics, 66(2),
463-506.
[35] Krishnamurthy, Arvind and Annette Vissing-Jorgensen, 2007, “The Aggregate Demand for Treasury
Debt,” working paper, Northwestern University.
[36] Xiong, Wei and Albert Kyle, 2001, “Contagion as a Wealth Effect,” Journal of Finance 56, 1401-1440.
[37] Lucas, Robert E., Jr, 1978, Asset Prices in an Exchange Economy, Econometrica, 46, 1429–1445,
[38] Mankiw, Gregory N, and Stephen Zeldes, 1991, The Consumption of Stockholders and Non-stockholders,
Journal of Financial Economics, 29, pp. 97-112.
[39] Mitchell, Mark, Lasse Pedersen, and Todd Pulvino, 2007, “Slow Moving Capital,” American Economic
Review Papers and Proceedings, forthcoming.
[40] Pastor, Lubos and Robert Stambaugh, 2003, Liquidity Risk and Expected Stock Returns, Journal of
Political Economy, 111, 642-685.
47
[41] Remolina, E. M., P. Kleiman, and D. Gruenstein (1997), Market Returns and Mutual Fund Flows,
Federal Reserve Bank of New York Economic Policy Review, 3 (2), pp. 33-43.
[42] Sadka, Ronnie, 2003, Momentum, Liquidity Risk, and Limits to Arbitrage, working paper, University
of Washington.
[43] Shleifer, Andrei and Robert W. Vishny, 1997, The Limits of Arbitrage, Journal of Finance 36, 35-55.
[44] Vayanos, Dimitri, 2004, Flight to Quality, Flight to Liquidity, and the Pricing of Risk, working paper,
LSE.
[45] Vissing-Jorgensen, Annette, 2002, Limited asset market participation and the elasticity of intertemporal
substitution, Journal of Political Economy, 110(4), 825-853.
[46] Warther, Vincent, 1995, Aggregate mutual fund flows and security returns, Journal of Financial Economics, 39, 209-235.
[47] Xiong, Wei, 2001, Convergence Trading with Wealth Effects: An Amplification Mechanism in Financial
Markets, Journal of Financial Economics 62, 247-292
48
A
ODE Solution
In this appendix, we detail the solution of the ODE that characterizes the equilibrium. We analyze our ODE based on
state variable y, i.e., the scaled households wealth. Since under our parameterization the equilibrium scaled specialists
wealth w/D = F (y) − y is a monotone transform of y, in the main text we plot F (·) against w/D to highlight the
effect of specialist wealth.
A.1
Derivation of µy and σy
We rewrite equation (5) which describes the wealth dynamics (budget constraint) of the household sector as:
dw h = θs dP + Dθs dt + rθ̂b dt + lDt dt − ρw hdt.
(23)
In this equation,
θs = αI αh (1 − λ)
wh
P
(24)
are the number of shares that the stock household owns, and
θ̂b D = w h − θs P
(25)
is the amount of funds that the stock and debt households together have invested in the riskless bond. αh and αI
are defined in the text and depends on whether the economy is constrained or not.
We apply Ito’s Lemma to P = DF (y) to find expressions for the drift and diffusion of dP . We can then substitute
back into equation (23) to find expressions for the drift and diffusion of dw h.
Now, we have defined w h = Dy. We apply Ito’s Lemma to this equation to arrive at a second expression for the
drift and diffusion of dw h. Matching the drift and diffusion terms from these two ways of writing dw h , we solve to
find µy and σy .
The result of this algebra is that:
σy = −
and,
µy =
A.2
1
1 − θs F 0
„
θ̂b
σ.
1 − θs F 0
«
1
θs + l + (r + σ2 − g)θ̂b − ρy + θs F 00 σy2 .
2
ODE
Because ct = Dt (1 + l − ρyt), we have
dct
ct
=
=
»
–
dDt
ρdy
ρ
dD
−
−
Covt dy,
Dt
1 + l − ρy
1 + l − ρy
D
„
«
„
«
ρ
ρσy
g−
(µy + σy σ) dt + σ −
dZt.
1 + l − ρy
1 + l − ρy
We also have
dRt =
»
–
„
«
dPt + Dt dt
F0
1 F 00 2
1
F0
F0
= g+
µy +
σy +
+
σy σ dt + σ +
σy dZt .
Pt
F
2 F
F
F
F
49
Substituting these expressions into (16) we obtain,
g+
F0
1 F 00 2
µy +
σ
F
2 F y
+
1
F0
γρ
+
σy σ = ρ + γg −
(µy + σy σ)
F
F
1 + l − ρy
„
«„
«
„
«2
ρ
F0
1
ρ
+γ σ −
σy
σ+
σy − γ(γ + 1) σ −
σy
1 + l − ρy
F
2
1 + l − ρy
which is equation (17) in Proposition 1.
Substituting for µy derived in (A.1), we find,
„ 0
«„
«“
”
F
γρ
1
1
+
θs + l + θ̂b (r − g) − ρy +
0
F
1 + l − ρy
1 − θs F
F
„
«„
«
1
1
1
γρ
+ F 00 σy2
+
θ
s
2
1 − θs F 0
F
1 + l − ρy
«„
«
„
0
ρ
F
σy
σ+
σy
= ρ + g(γ − 1) + γ σ −
1 + l − ρy
F
„
«2
1
ρ
− γ(γ + 1) σ −
σy
2
1 + l − ρy
where,
r
=
θs + l + (r − g) θ̂b − ρy +
ργ
ρ + gγ −
1 + l − ρy
1 − θs F 0
!2
γ (γ + 1) σ2
ρθ̂b
1
−
1+
2
1 + l − ρy 1 − θs F 0
We define a function, G(y) ≡
1
;
1−θs F 0
θ̂b2
00
σ2
2 θs F (1−θs F 0 )2
with this definition, we can write G0 = θs G2 F 00 , and
σy = −
θ̂b
σ = −θ̂b σG.
1 − θs F 0
Therefore we have
G0
(θ̂b σ)2
2
G
„
1
γρh
+
θs F
1 + l − ρy
and
r=
ρ + gγ −
«
1
F
”
0 “
1
„
« 2 y − Gθ̂b
1 2
ρ
1
+
l
−
ρy
+
ρG
θ̂
b
A
+ γσ 1 +
θ̂b G @
− (1 + γ)
2
1 + l − ρy
θs F
1 + l − ρy
„
«“
”
G−1
γρh
+
G
θs + l + θ̂b (r − g) − ρy
−
θs F
1 + l − ρy
ργG
1+l−ρy
“
= ρ + g(γ − 1) −
θs + l − gθ̂b − ρy +
1+
“
ργGθ̂b
1+l−ρy
We combine these two pieces, using the relation, θ̂b G−1
+
θs F
final expression of the ODE:
„
«
(θ̂b σ)2 G
1 + l + ρy(γ − 1)
G0
2 θs F 1 + l − ρy + ργGθ̂b
=
σ2 0 2
G θ̂b
2
γρh G
1+l−ρy
”
”
−
γ(γ+1)σ 2
2
“
1+
θ̂b
= − y−G
+
θs F
ρθ̂b G
1+l−ρy
”2
.
1+l−ρy+γρh Gθ̂b
,
1+l−ρy
and arrive at a
#
!
"
1
γ (1 − γ) σ2
ρGθ̂b
y − Gθ̂b 1 + l − ρy − ρGθ̂b
ρ + g(γ − 1) −
+
1+
F
2
1 + l − ρy
θs F
1 + l − ρy + ργGθ̂b
„
«
(1 + l − ρy) (G − 1)
θs + l + θ̂b (g(γ − 1) + ρ) − ρy
−
+ γρh G
.
θs F
1 + l − ρy + ργGθ̂b
50
(26)
The expressions for the bond holding θ̂b and stock holding θs depend on whether the economy is constrained or
not. In the unconstrained region, as shown in Section 3.3, αh = 1, and αI =
θs =
(1−λ)y
F −λy ,
θbb = y −
−y
and θbb = λy FF−λy
. In the constrained region αh =
m
1+m F .
m(F −y)
(1−λ)y ,
F
.
F −λy
αI =
Utilizing (25) and (24), we have
1
F
1+m F −y ,
m
1+m ,
and
m
c
1−λ+m F (y ),
and
therefore θs =
Finally, as illustrated in Section 3.3, the cutoff for the constraint satisfies yc =
the economy is in the unconstrained region if 0 < y ≤ yc .
A.3
Boundary conditions and technical parameter restriction
As described in the text an upper boundary for y occurs when yb =
1+l
ρ .
The boundary condition is that F (yb ) = yb .
In Appendix B, we show that the condition F (yb ) − yb = 0 is required in order to satisfy the transversality condition
in the specialist’s budget equation; with this transversality condition we are able to show that the equilibrium is
well-defined.
` ´
Also, a straightforward calculation yields that F 0 yb = 1 if F (yb ) = yb . This result also ensures that the mapping
from the scaled household’s wealth y to the scaled specialist wealth w/D = F (y) − y is strictly decreasing in the
` ´
scaled household’s wealth y (this monotone relation clearly fails if F yb > yb .) As a result, it is equivalent to model
either agent’s wealth as our state variable.
A lower boundary condition occurs when y → 0. This case corresponds to one where specialists hold the entire
0 (0)
→ FF (0)
. Plugging this result
financial wealth of the economy. Using L’Hopital’s rule, it is easy to check that G−1
θs F
into (26), and noting that both θs and θbb go to zero as y goes to zero, we obtain,
F (0) =
1 + F 0 (0) l
ρ + g(γ − 1) +
γ(1−γ)σ 2
2
−
lγρ
1+l
.
(27)
When l = 0, one can check that F (0) is the equilibrium Price/Dividend ratio for the economy with the specialists as
the representative agent. However because in our model the growth of the household sector affects the pricing kernel,
this boundary P/D ratio F (0) also depends on the household’s labor income l. As in the case where l = 0, for the
P/D ratio to be well defined we require that parameters satisfy,
ρ + g(γ − 1) +
A.4
γ (1 − γ) σ2
lγρ
−
> 0.
2
1+l
(28)
Numerical Method
In our ODE (26) both boundaries are singular, causing difficulties in directly applying the built-in ODE solver ode15s
in M atlab. To overcome this issue, we use a similar modification as in He and Krishnamurthy (2006). Specifically, we
`
` ´
´
`
´
approximate the upper-end boundary yb , F yb = yb by yb − η, y b − η (where η is sufficiently small), and adopt
a “forward-shooting and line-connecting” method for the lower–end boundary. Take a small > 0 and call Fe as the
attempted solution. For each trial φ ≡ Fe 0 (), we set Fe0 (0) = φ, solve Fe (0) based on (27), and let Fe () = Fe (0) + φ.
51
Since
“
, Fe ()
”
is away from the singularity, by trying different φ’s we apply the standard shooting method to
`
´
obtain the desired solution F that connects at yb − η, y b − η . For y < , we simply approximate the solution by a
ˆ
˜
line connecting (0, F (0)) and (, F ()). In other words, we solve F on , y b with a smooth pasting condition for
F 0 () =
F ()−F (0)
` ´
and a value matching condition for F yb = yb .
We use = 0.1 and η = 0.001 which give ODE errors bounded by 3 × 10−5 for y > . Different ’s and η’s
deliver almost identical solutions for y > 1. Because we are mainly interested in the solution behavior near yc (which
takes a value of 14 even in the m = 1 case) and onwards, our main calibration results are free of the approximation
errors caused by the choice of and η. Finally we find that, in fact, these errors are at the same magnitude as those
generated by the capital constraint around yc (3.5 × 10−5 ).
B
Verification of optimality
In the section we take the equilibrium Price/Dividend ratio F (y) as given, and verify that the specialist’s consumption
policy c = Dt (1 + l − yt ) is optimal subject to his budget constraint.
Our argument is a variant of the standard one: it uses the strict concavity of u (·) and the specialist’s budget
constraint to show that the specialist’s Euler equation is necessary and sufficent for the optimality of his consumption
plan.
Specifically, fixing t = 0 and the starting state (y0 , D0 ), define the pricing kernel as
ξt ≡ e−ρtc−γ
= e−ρt Dt−γ (1 + l − ρyt)−γ .
t
R∞
Consider another consumption profile b
c which satisfies the budget constraint E 0 b
ctξt dt ≤ ξ0 D0 (F0 − y0 ) (recall that
the specialist’s wealth is D0 (F0 − y0 ); here we require that the specialist’s feasible trading strategies be well-behaved,
e.g., his wealth process remains non-negative). Then we have
Z ∞
Z ∞
Z ∞
E
e−ρt u (ct ) dt ≥ E
e−ρt u (b
ct ) dt + E
e−ρt u0 (ct ) (ct − b
ct) dt
0
0
0
Z ∞
Z ∞
Z ∞
ξt b
ctdt.
ξt ctdt − E
= E
e−ρt u (b
ct ) dt + E
0
0
0
If the specialist’s budget equation holds in equality for the equilibrium consumption process c, i.e., if
Z ∞
E
ξt ct dt = ξ0 D0 (F0 − y0 ) ,
0
then the result follows. Somewhat surprisingly, for our model this seemingly obvious claim requires an involved
argument because of the singularity at yb = 1+l
.
ρ
One can easily check that, for ∀T > 0, we have
Z T
Z
ξ0 D0 (F0 − y0 ) =
ctξt dt +
0
0
T
σ (Dt, yt ) dZt + ξT DT (FT − yT ) ,
(29)
where σ (Dt , yt ) corresponds to the specialist’s equilibrium trading strategy (which involves terms such as (1 + l − ρy)−γ−1
and is NOT uniformly bounded as y → yb ). Our goal in the following steps is to show that in expectation, the latter
two terms vanishes when T → ∞.
52
Step 1: Limiting Behavior of y at yb
The critical observation regarding the evolution of y is that when y
approaches yb , it approximately follows a Bessel process with a dimension δ = γ + 2 > 2. (Given a δ-dimensional
qP
δ
2
Brownian motion Z, a Bessel process with a dimension δ is the evolution of kZk =
i=1 Zi , which is the Euclidean
distance between Z and the origin.) According to standard results on Bessel processes, yb is an entrance-no-exit point,
and is not reachable if the starting value y0 < yb (if δ > 2). Intuitively, when y is close to yb , the dominating part of
µy is proportional to
1
y−y b
< 0, while the volatility σy is bounded— therefore a drift that diverges to negative infinity
keeps y away from the singular point yb . This result implies that our economy never hits yb .
To show that for y close to yb , y’s evolution can be approximated by a Bessel Process, one can easily check that
when y → yb ,
(γ + 1) σ2 ρh θ̂b G
(γ + 1) σ2 ρh θ̂b2 G2
r'−
, µy ' −
, σy = −Gσθ̂b ;
h
2
1+l−ρ y
2
1 + l − ρh y
and therefore
(γ + 1) σ2 ρθ̂b2 G2
dt − Gσθ̂b dZ t .
dy = −
2
1 + l − ρy
`
´
1
Utilizing the result F 0 yb = 1 established in Section A.3, we know that when y → yb , θ̂b ' F − θs y ' 1+m
yb =
1 1+l
, and G ' 1 + m. Let
1+m ρ
xt = 1 + l − ρyt ;
then it is easy to show that q =
x
Gσ θ̂b ρ
=
x
σ(1+l)
dq = −
evolves approximately according to
1
Gσθ̂b
dy =
(γ + 1)
dt + dZ t ,
2q
which is just a standard Bessel process with a dimension δ = γ + 2. Therefore, x is also a scaled version of a Bessel
process, and can never reach 0 (or, y cannot reach yb ). In the following analysis, we focus on the limiting behavior
of x.
Step 2: Localization Note that in (29), due to the singularity at x = 0 (or, y = yb ), both the local martingale
RT
part 0 σ (Dt, yt ) dZt and the terminal wealth part ξT DT (FT − yT ) are not well-behaved. To show our claim, we
have to localize our economy, i.e., stop the economy once y is sufficiently close to yb (or, once D is sufficiently close
to 0). Specifically, we define

ff
1
1
Tn = inf t : xt =
or Dt = h
n
n
where h is a positive constant (as we will see, the choice of h, which is around 1, gives some flexibility for γ other
than 2). Since y and x have a one-to-one relation (x = 1 + l − ρy), for simplicity we localize x instead.
RT
Clearly this localization technique ensures that the local martingale part 0 n σ (Dt, yt ) dZt is a martingale (one
can check that σ (Dt , yt ) is continuous in Dt and yt, in turn Dt and xt ; therefore σ (Dt , yt ) is locally bounded). As
Tn → ∞ when n → ∞, for our claim we need to show
lim E [ξTn DTn (FTn − yTn )] = 0
n→∞
We substitute from the definition of ξ:
h
i
h
i
−ρTn h(γ−1) −γ
E e−ρTn DT1−γ
x−γ
n
xTn (F (yTn ) − yTn ) .
Tn (F (yTn ) − yTn ) ≤ E e
n
Since the analysis will be obvious if x−γ (F (y) − y) is uniformly bounded (notice here x = 1 + l−
to
“ ρy), it ”is sufficient
“
”
` b´
` b´
b
0
b
b
1
1
1
consider xTn = n . Because F y = y and F y = 1, by Taylor expansion we know that F y − nρ − y − nρ
53
can be written as ψ (n)
n → ∞,
1
n
when n is sufficiently large, and ψ (n) → 0 as n → ∞. Therefore we have to show that, as
and a sufficient condition is that,
h
i
E e−ρTn n(γ−1)(1+h) ψ (n) → 0
h
i
E e−ρTn n(γ−1)(1+h) → K
where K is bounded.
We apply existing analytical results in the literature to show our claim. To do so, we have to separate our two
state variables. We define

ff

ff
1
1
TnD = inf t : Dt = h , Tnx = inf t : xt =
.
n
n
h
i
h
i
ˆ
˜
D
x
We want to bound E e−ρTn by the sum of E e−ρTn and E e−ρTn ; note that they are Laplace transforms of the
first-hitting time distribution of a GBM and Bessel processes, respectively. The Laplace transform of Tn is simply
Z ∞
h
i Z ∞
E e−ρTn =
e−ρT dF (T ) = ρ
e−ρT F (T ) dT,
0
0
where the bold F denotes the distribution function of Tn . The similar relation also holds for TnD or Tnx . Denote FD (·)
(or Fx (·)) as the distribution function for TnD (or Tnx ), and notice that
“
”
“
”
1 − F (T ) = Pr (Tn > T ) = Pr TnD > T, Tnx > T > Pr TnD > T Pr (Tnx > T )
=
1 − FD (T ) − Fx (T ) + FD (T ) Fx (T ) ,
because 1{T D >T } and 1{T x >T } are positively correlated (both take the value 1 when Z is high).15
n
n
F (T ) < FD (T ) + Fx (T ), or
h
i
h
i
h
i
D
x
E e−ρTn n(γ−1)(1+h) < E e−ρTn n(γ−1)(1+h) + E e−ρTn n(γ−1)(1+h)
Therefore
Using the standard result of the Laplace transform of the first-hitting time distribution for a GBM process, we can
h
i
D
easily verify that as n → ∞, the first term E e−ρTn n(γ−1)(1+h) vanishes under our parameters when h = 0.9 (in
fact, this relates to the parameter restriction for a standard GBM/CRRA economy).
Step 3: Regulated Bessel Process
The challenging task is the second term. Notice that our economy (i.e.,
evolution of x) differs from the evolution of a Bessel process when x is far away from 0; therefore an extra care needs
to be taken. We consider a regulated Bessel process which is reflected at some positive constant x. Intuitively, by
h
i
x
doing so, we are putting an upper bound for E e−ρTn , as the reflection makes xt to hit n1 more likely (therefore, a
larger Fx ). Also, for a sufficiently small x > 0, when x ∈ (0, x], x can be approximated by a Bessel process with a
dimension γ + 2 − ε. Therefore, Fx must be bounded by the first-hitting time distribution of a Bessel process with a
dimension δ, where δ takes value from γ + 2 − to γ + 2, where is sufficiently small. Finally, note that by considering
15 Technically, using the technique of Malliavian derivatives, we can show that both x and D have positive
s
s
diffusions in the martingale representations for all s. Then, the running minimum xT = min {xt : 0 < t < T } and
D T = min {Dt : 0 < t < T } have positive loadings always on the martingale representations (using the technique in
Methods of Mathematical Finance, Karatzas and Shreve (1998), Page 367). The same technique can be applied to
1{T x >T } = 1{x >T } and 1{T D >T } = 1{D >T } , as an indicator function can be approximated by a sequence of
n
T
n
T
differentiable increasing functions.
54
a Bessel process we are neglecting certain drift for x. However, one can easily check that when x is close to 0, the
adjustment term for µy is − 1+l
γσ2 < 0. This implies that we are neglecting a positive drift for x—which potentially
ρ
makes hitting less likely—thereby yielding an upper-bound estimate.
We have the following Lemma from the Bessel process.
Lemma 1 Consider a Bessel process ˘x with δ >¯2 which is reflected at x > 0. Let ν =
we consider the hitting time Tnx = inf t : xt = n1 . Then we have
h
i
x
E e−ρTn ∝ n−2ν as n → ∞
δ
2
− 1. Starting from x0 ≤ x,
Proof. Due to the standard results in Bessel process and the Laplace transform of the hitting time (e.g., see Borodin
and Salminen (1996), Chapter 2), we have
h
i ϕ (x )
x
0
E e−ρTn = ` 1 ´ ,
ϕ n
where
ϕ (z) = c1 z −ν Iv
“p
”
“p
”
2ρz + c2 z −ν Kv
2ρz ,
and Iv (·) (and Kv (·)) is modified Bessel function of the first (and second) kind of order v. Because R is a reflecting
barrier, the boundary condition is
ϕ0 (x) = 0,
which pins down the constants c1 and c2 (up to a constant multiplication; notice that this does not affect the value
h
i
h
i
`√
´
x
x
of E e−ρTn ). Therefore the growth rate of E e−ρTn is determined by nν Kv
2ρn−1 as Kv dominates Iv near
0. Since Kv (x) has a growth rate x−ν when x → 0, the result is established.
h For any
i y0 , redefine starting point as x0 = min (1 + l − y0 , x); clearly this leads to an upper-bound estimate for
x
E e−ρTn . However, since for all δ ∈ [γ + 2 − , γ + 2], the above Lemma tells us that for any ε ∈ [0, ], when n → ∞,
h
i
x
n(γ−1)(1+h) E e−ρTn ∝ n(γ−1)(1+h) n−2ν = n(γ−1)(1+h)−γ+ε → 0
uniformly if γ = 2 and h = 0.9 (and for some sufficiently small > 0). Therefore we obtain our desirable result.
Finally ct ξt > 0 implies that
R∞
0
ct ξtdt converges monotonically, and therefore the specialist’s budget equation
T →∞
RT
C
Government Debt
lim E
0
ξt ct dt = ξ0 D0 (F0 − y0 ) holds for all stopping times that converge to infinity. Q.E.D.
Goverment debt outstanding at t is Bt = max (0, bDt (yt − y∗ )). In the region of yt > y∗ , Bt = bwth − bDt y∗ , and the
tax paid by households is
“ ”
τt = bwth rt dt − bDt y∗ rt dt − bd wth + by∗ dDt .
Substituting into the household’s budget equation, dw h = θs dP + Dθs dt + rt θb dt + lDdt − ρwthdt − τt , we find:
σy
=
µy
=
θbb − b (y − y∗ )
σ
1 − b − θs F 0
„
«
”
`
´“
1
2
bb − b (y − y∗ ) − ρy + 1 θs F 00 σy2 .
θ
+
l
+
r
+
σ
−
g
θ
s
1 − b − θs F 0
2
−
55
The ODE remains the same as before (equation (17)); after substituting for µy and σy , and collecting the terms, we
arrive at the final ODE:
0
1
∗ 2 2
∗
(
θ̂
−
b
(y
−
y
))
σ
G
(1
+
l)
(1
−
b)
+
ρy(γ
−
1)
(1
−
b)
+
ργby
b
@
A
“
”
G0
2
θs F
1 + l − ρy + ργG θ̂b − b (y − y∗ )
=
1
+
ρ + g(γ − 1) −
F
“
”1
“
”
“
” 3
0
2
∗
∗
∗
∗
2
ρG
θ̂
−
b
(y
−
y
)
(1
−
b)
y
−
(1
−
b)
G
θ̂
−
b
(y
−
y
)
+
by
1
+
l
−
ρy
−
ρG
θ̂
−
b
(y
−
y
)
b
b
b
γ (1 − γ) σ @
A
4
“
”5
1+
2
1 + l − ρy
θs F
1 + l − ρy + ργG θ̂b − b (y − y∗ )
“
”
„
« θs + l + θ̂b − b (y − y∗ ) (g(γ − 1) + ρ) − ρy
(1 + l − ρy) ((1 − b) G − 1)
“
”
−
+ γρG
,
θs F
1 + l − ρy + ργG θ̂b − b (y − y∗ )
∗
∗
1
where G(y) ≡ 1−b−θ
0 again. Clearly this is for y > y ; when y < y , we can simply set b = 0.
sF
Now we derive the boundary conditions. For y = 0, the same boundary condition (20) as without government
debt applies. When y = 1+l
ρ , the specialist’s wealth is zero, and we must have
«
1+l
− y∗ ,
ρ
“
”
∗
where the LHS is the total wealth in this economy. This implies that F 1+l
= 1+l
ρ
ρ (1 − b) + by .
c
∗
Finally, in the numerical solutions we focus on the case where y > y , i.e., the government issue some public
debt
the` capital constraint
is binding. Therefore, the capital constraint binds when (1 − λ) w h = mw =
` before
´´
h
h
∗
m P − w + b w − Dy , and this implies,
1+l
=F
ρ
„
yc =
1+l
ρ
«
+b
„
m (F c − by∗ )
.
1 − λ + m (1 − b)
56